File size: 2,803 Bytes
a4b0c5b | 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 | from __future__ import annotations
import importlib.util
from pathlib import Path
import pytest
pytest.importorskip("torch")
pytest.importorskip("onnxruntime")
pytest.importorskip("transformers")
from vons.models import ModelConfig, build_model
_SPEC = importlib.util.spec_from_file_location(
"verify_onnx_bundle", Path(__file__).parents[1] / "tools/verify_onnx_bundle.py"
)
assert _SPEC is not None and _SPEC.loader is not None
_MODULE = importlib.util.module_from_spec(_SPEC)
_SPEC.loader.exec_module(_MODULE)
_model_config_from_checkpoint = _MODULE._model_config_from_checkpoint
def test_token_conditioned_checkpoint_config_reconstructs_model_state() -> None:
expected = ModelConfig(
hidden_size=256,
latent_size=64,
diffusion_steps=64,
inference_steps=8,
abstain_threshold=0.55,
diffusion_conditioning="token_cross_attention",
attention_heads=4,
)
trained_head = build_model("diffusion", option_count=32, config=expected)
checkpoint = {
"backend": "diffusion",
"hidden_size": 256,
"config": {
"diffusion_conditioning": "token_cross_attention",
"latent_size": 64,
"diffusion_steps": 64,
"inference_steps": 8,
"abstain_threshold": 0.55,
"attention_heads": 4,
},
}
reconstructed = _model_config_from_checkpoint(checkpoint)
verifier_head = build_model("diffusion", option_count=32, config=reconstructed)
verifier_head.load_state_dict(trained_head.state_dict())
assert reconstructed == expected
def test_token_conditioned_checkpoint_uses_training_defaults() -> None:
checkpoint = {
"backend": "diffusion",
"hidden_size": 256,
"config": {"diffusion_conditioning": "token_cross_attention"},
}
reconstructed = _model_config_from_checkpoint(checkpoint)
assert reconstructed.latent_size == 64
assert reconstructed.diffusion_steps == 64
assert reconstructed.inference_steps == 8
assert reconstructed.diffusion_conditioning == "token_cross_attention"
def test_pooled_checkpoint_without_schedule_fields_uses_export_defaults() -> None:
expected = ModelConfig(
hidden_size=256,
latent_size=32,
diffusion_steps=32,
inference_steps=4,
diffusion_conditioning="pooled",
)
trained_head = build_model("diffusion", option_count=32, config=expected)
checkpoint = {
"backend": "diffusion",
"hidden_size": 256,
"config": {},
}
reconstructed = _model_config_from_checkpoint(checkpoint)
verifier_head = build_model("diffusion", option_count=32, config=reconstructed)
verifier_head.load_state_dict(trained_head.state_dict())
assert reconstructed == expected
|