Repository navigation
Expand file tree
/
Copy pathspeech-catalog.test.ts
More file actions
157 lines (150 loc) · 6.18 KB
/
Copy pathspeech-catalog.test.ts
File metadata and controls
157 lines (150 loc) · 6.18 KB
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
// @effect-diagnostics nodeBuiltinImport:off - compares the generated bundle directly with its checked-in bytes.
import * as NodeFS from "node:fs";
import { catalog } from "@handy-computer/models";
import { recommendSpeechModels } from "@t3tools/shared/speech";
import policy from "../../../packages/shared/src/speechRecommendationPolicy.json" with { type: "json" };
import { expect, it } from "vite-plus/test";
import manifest from "../package.json" with { type: "json" };
import { generateSpeechCatalog, renderSpeechCatalog } from "./speech-catalog.ts";
const generate = (models = catalog.models) =>
generateSpeechCatalog(
{ ...catalog.data, models },
manifest.devDependencies["@handy-computer/models"],
manifest.dependencies["transcribe-cpp"],
);
it("reproduces the bundled registry byte for byte without upstream requests", async () => {
const { fastCpuMaxWaitSeconds, ...methodology } = catalog.methodology;
expect(policy).toEqual({ ...methodology, fastMaxWaitSeconds: fastCpuMaxWaitSeconds });
expect(NodeFS.readFileSync(new URL("../src/speech/catalog.json", import.meta.url), "utf8")).toBe(
await renderSpeechCatalog(
catalog.data,
manifest.devDependencies["@handy-computer/models"],
manifest.dependencies["transcribe-cpp"],
),
);
for (const model of generate().models) {
const source = catalog.models.find((entry) => entry.repository === model.id)!;
expect(model).toMatchObject({
revision: source.revision,
name: source.name,
...source.files[source.quant],
supportsStreaming: source.capabilities.streaming,
supportsLanguageDetection: source.capabilities.languageDetection,
supportsTranslation: source.capabilities.translate,
});
}
const parakeet = catalog.models[0]!;
const withoutFrenchOrSpeed = generateSpeechCatalog(
{
...catalog.data,
models: [{ ...parakeet, languages: ["en", "fr"] }],
results: { [parakeet.id]: { accuracy: { en: { error: 3.99 } }, performance: {} } },
},
"0.2.3",
"0.4.1",
);
expect(withoutFrenchOrSpeed.models[0]?.benchmark).toEqual({
accuracy: { en: { grade: "A-", errorPercent: 3.99 } },
secondsPer30s: { cpu: null, accelerated: null },
});
expect(generate().models[0]?.benchmark.secondsPer30s.accelerated).toBeCloseTo(1.19, 2);
for (const performance of [{ cpu: 2 }, { accelerated: 10 }, { cpu: 2, accelerated: 10 }]) {
const result = generateSpeechCatalog(
{
...catalog.data,
models: [parakeet],
results: { [parakeet.id]: { accuracy: { en: { error: 3.99 } }, performance } },
},
"0.2.3",
"0.4.1",
);
expect(result.models[0]?.benchmark.secondsPer30s).toEqual({
cpu: performance.cpu ? 30 / performance.cpu : null,
accelerated: performance.accelerated ? 30 / performance.accelerated : null,
});
}
});
it("preserves Handy's accelerated balanced pick after filtering T3's offered candidates", () => {
const offered = generate().models;
const sources = catalog.models.filter((model) =>
offered.some(({ id }) => id === model.repository),
);
for (const languages of [["en"], ["de"], ["ja"], ["en", "de"], ["en", "fr"], ["hi", "mr"]]) {
const expected = catalog
.recommendModels(sources, languages)
.find((pick) => pick.status !== "unbenchmarked" && pick.roles.includes("best"));
expect(
recommendSpeechModels(offered, languages, "accelerated").keys().next().value,
languages.join(","),
).toBe(expected?.model.repository);
}
const excluded = { ...catalog.models[0]!, license: "cc-by-nc-4.0" };
const allowed = { ...catalog.models[1]!, languages: ["en"] };
const filtered = generate([excluded, allowed]).models;
expect([...recommendSpeechModels(filtered, ["en"], "accelerated").keys()]).toEqual([
allowed.repository,
]);
});
it("filters restricted licenses, unsupported roles, engines, architectures and files", () => {
const model = catalog.models[0]!;
for (const override of [
{ role: "langid" as const },
{ role: "vad" as const },
{ role: "diarize" as const },
{ license: "cc-by-nc-4.0" },
{ license: "CC-BY-NC-SA-4.0" },
{ license: "cc-by-nc-nd-3.0" },
{ license: "cc-by-nc" },
{ license: "CC BY-NC 4.0" },
{ license: "custom_NC" },
{ minEngine: "0.4.2" },
{ minEngine: "0.5.0" },
{ family: "future_arch" },
...[
"nested/model.gguf",
"nested\\model.gguf",
"../model.gguf",
"model.bin",
"model.safetensors",
].map((filename) => ({ files: { [model.quant]: { ...model.files[model.quant]!, filename } } })),
]) {
expect(generate([{ ...model, ...override }]).models).toEqual([]);
}
for (const license of ["other", "apache-2.0", "cc-by-4.0", "cc-by-sa-4.0", "unlicense"]) {
expect(generate([{ ...model, license, minEngine: "0.4.1" }]).models).toHaveLength(1);
}
const moonshine = catalog.models.find((entry) => entry.id === "moonshine-tiny-ja")!;
for (const [id, allowed] of [
["moonshine-tiny-ja", false],
["moonshine-base-vi", false],
["moonshine-base-es", false],
["moonshine-base", true],
["moonshine-streaming-tiny-ja", true],
] as const) {
const entry = { ...moonshine, id, repository: `handy-computer/${id}-gguf`, license: "mit" };
expect(generate([entry]).models, id).toHaveLength(allowed ? 1 : 0);
}
});
it("rejects missing defaults, unpinned or corrupt metadata, collisions, and unreviewed engine/schema changes", () => {
const model = catalog.models[0]!;
const file = model.files[model.quant]!;
for (const override of [
{ files: {} },
{ revision: "main" },
{ repository: "../model" },
{ languages: [] },
...[{ sha256: "bad" }, { size: 0 }, { size: 1.5 }].map((entry) => ({
files: { [model.quant]: { ...file, ...entry } },
})),
]) {
expect(() => generate([{ ...model, ...override }])).toThrow();
}
expect(() => generate([model, model])).toThrow(/Duplicate/);
expect(() => generate([model, { ...model, repository: "handy-computer/another" }])).toThrow(
/Duplicate/,
);
expect(() => generateSpeechCatalog(catalog.data, "0.2.3", "0.4.2")).toThrow(/loader support/);
expect(() =>
generateSpeechCatalog({ ...catalog.data, schema: 3 as 2 }, "0.2.3", "0.4.1"),
).toThrow(/schema/);
});