from __future__ import annotations
import copy
import json
import pathlib
import jsonschema
ROOT = pathlib.Path(__file__).resolve().parents[1]
def valid_vectors() -> dict:
return {
"schema": "denoize-runtime-model-numerical-vectors-v1",
"profile_id": "fp32",
"cases": [
{
"id": "identity",
"inputs": [
{
"name": "input",
"element_type": "float32",
"shape": [1, 4],
"values": [-0.5, 0.0, 0.25, 0.75],
}
],
"outputs": [
{
"name": "output",
"element_type": "float32",
"shape": [1, 4],
"values": [-0.5, 0.0, 0.25, 0.75],
}
],
"tolerance": {"absolute": 0.000001, "relative": 0.000001},
}
],
}
def valid_manifest() -> dict:
digest = "0" * 64
resources = {
"max_session_memory_bytes": 67108879,
"max_worker_memory_bytes": 4096,
"max_gpu_session_memory_bytes": 0,
"max_gpu_worker_memory_bytes": 0,
"accelerators": ["cpu"],
}
tensor = lambda name: {
"name": name,
"role": "audio",
"element_type": "float32",
"axes": [
{"name": "batch", "kind": "batch", "fixed": 1},
{"name": "samples", "kind": "sample", "fixed": None},
],
"optional": False,
"state_id": None,
}
return {
"schema": "denoize-runtime-model-package-v2",
"format_version": 2,
"package_id": "example.identity-v2",
"package_revision": "1",
"signing_key_id": "0123456789ABCDEF",
"runtime": {
"kind": "onnx-audio-graph-v2",
"sample_rate_hz": 16000,
"mode": "finite-and-streaming",
},
"frontend": {
"normalization": "pcm-f32-minus-one-to-one-v1",
"resampling": "bandlimited-waveform-v1",
"duration": "preserve-input-frames-v1",
"channels": {
"policy": "independent-mono",
"roles": [],
"geometry": None,
},
},
"tensors": {"inputs": [tensor("input")], "outputs": [tensor("output")]},
"state_pairs": [],
"latency": {
"frame_samples": 4,
"hop_samples": 4,
"left_context_samples": 0,
"right_context_samples": 0,
"lookahead_samples": 0,
"algorithmic_latency_samples": 0,
"flush_samples": 0,
},
"components": [
{
"id": "model-fp32",
"kind": "onnx-model",
"file": {"filename": "model.onnx", "size_bytes": 5, "sha256": digest},
},
{
"id": "license",
"kind": "license-notice",
"file": {"filename": "LICENSE.txt", "size_bytes": 7, "sha256": digest},
},
{
"id": "provenance",
"kind": "provenance-json",
"file": {"filename": "provenance.json", "size_bytes": 2, "sha256": digest},
},
{
"id": "vectors-fp32",
"kind": "numerical-vectors-json",
"file": {"filename": "vectors.json", "size_bytes": 32, "sha256": digest},
},
],
"precision_profiles": [
{
"id": "fp32",
"element_type": "float32",
"model_component": "model-fp32",
"numerical_vectors_component": "vectors-fp32",
"resources": resources,
}
],
"default_precision_profile": "fp32",
"license": {"spdx": "MIT", "notice_component": "license"},
"provenance": {
"component": "provenance",
"source_repository": "https://example.invalid/source",
"source_revision": "0123456789abcdef",
"source_sha256": digest,
"source_license_spdx": "MIT",
"checkpoint_source": "https://example.invalid/checkpoint",
"checkpoint_sha256": "1" * 64,
"checkpoint_license_spdx": "MIT",
"conversion_tool": "example-converter",
"conversion_revision": "1",
"training_datasets": [
{
"id": "synthetic",
"source": "urn:denoize:test:synthetic",
"revision": "1",
"sha256": "2" * 64,
"license_spdx": "CC0-1.0",
}
],
},
}
def main() -> None:
manifest_schema = json.loads(
(ROOT / "schemas/denoize-runtime-model-package-v2.schema.json").read_text(
encoding="utf-8"
)
)
vectors_schema = json.loads(
(
ROOT
/ "schemas/denoize-runtime-model-numerical-vectors-v1.schema.json"
).read_text(encoding="utf-8")
)
jsonschema.Draft202012Validator.check_schema(manifest_schema)
jsonschema.Draft202012Validator.check_schema(vectors_schema)
manifest_validator = jsonschema.Draft202012Validator(manifest_schema)
vectors_validator = jsonschema.Draft202012Validator(vectors_schema)
manifest = valid_manifest()
vectors = valid_vectors()
manifest_validator.validate(manifest)
vectors_validator.validate(vectors)
gpu_resources = copy.deepcopy(manifest)
gpu_resources["precision_profiles"][0]["resources"]["accelerators"] = ["cuda"]
manifest_validator.validate(gpu_resources)
semantic = copy.deepcopy(manifest)
semantic["tensors"]["inputs"].append(
{
"name": "query",
"role": "query",
"element_type": "float32",
"axes": [
{"name": "batch", "kind": "batch", "fixed": 1},
{"name": "classes", "kind": "feature", "fixed": 2},
],
"optional": False,
"state_id": None,
}
)
residual = copy.deepcopy(semantic["tensors"]["outputs"][0])
residual["name"] = "residual"
residual["role"] = "residual"
semantic["tensors"]["outputs"].append(residual)
manifest_validator.validate(semantic)
open_text = copy.deepcopy(semantic)
open_text["tensors"]["inputs"][1]["role"] = "natural-language"
assert not manifest_validator.is_valid(open_text)
unknown = copy.deepcopy(manifest)
unknown["command"] = "python converter.py"
assert not manifest_validator.is_valid(unknown)
scripted = copy.deepcopy(manifest)
scripted["components"][0]["kind"] = "script"
assert not manifest_validator.is_valid(scripted)
unsafe_name = copy.deepcopy(manifest)
unsafe_name["components"][0]["file"]["filename"] = "../model.onnx"
assert not manifest_validator.is_valid(unsafe_name)
private_source = copy.deepcopy(manifest)
private_source["provenance"]["source_repository"] = "/home/user/source"
assert not manifest_validator.is_valid(private_source)
secret_source = copy.deepcopy(manifest)
secret_source["provenance"]["checkpoint_source"] = (
"https://user:secret@example.invalid/checkpoint?token=secret"
)
assert not manifest_validator.is_valid(secret_source)
fragment_source = copy.deepcopy(manifest)
fragment_source["provenance"]["training_datasets"][0]["source"] = (
"urn:denoize:dataset#private-fragment"
)
assert not manifest_validator.is_valid(fragment_source)
oversized_license = copy.deepcopy(manifest)
oversized_license["components"][1]["file"]["size_bytes"] = 16777217
assert not manifest_validator.is_valid(oversized_license)
optional_output = copy.deepcopy(manifest)
optional_output["tensors"]["outputs"][0]["optional"] = True
assert not manifest_validator.is_valid(optional_output)
unknown_vector = copy.deepcopy(vectors)
unknown_vector["cases"][0]["outputs"][0]["path"] = "/tmp/output"
assert not vectors_validator.is_valid(unknown_vector)
unbounded_tolerance = copy.deepcopy(vectors)
unbounded_tolerance["cases"][0]["tolerance"]["absolute"] = 0.010001
assert not vectors_validator.is_valid(unbounded_tolerance)
float_overflow = copy.deepcopy(vectors)
float_overflow["cases"][0]["inputs"][0]["values"][0] = 3.5e38
assert not vectors_validator.is_valid(float_overflow)
fractional_integer = copy.deepcopy(vectors)
fractional_integer["cases"][0]["inputs"][0]["element_type"] = "int64"
fractional_integer["cases"][0]["inputs"][0]["values"][0] = 0.5
assert not vectors_validator.is_valid(fractional_integer)
if __name__ == "__main__":
main()