@flaxia/node 0.6.1 → 0.7.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +2 -0
- package/dist/assets/ai-inference.js +83 -183
- package/dist/assets/ai-model-registry.js +127 -0
- package/dist/assets/cc-extract.js +5166 -0
- package/dist/assets/container.js +59 -43
- package/dist/assets/decision.js +149 -0
- package/dist/assets/embed-gemma.js +105 -0
- package/dist/assets/image-process.js +78 -70
- package/dist/assets/nudenet.js +2 -2
- package/dist/assets/{VectorStoreEngine.js → sharedEngine.js} +67 -41
- package/dist/assets/transformers.web.js +2 -2
- package/dist/assets/vector-embed.js +148 -25
- package/dist/assets/vector-query.js +4 -4
- package/dist/assets/vector-store.js +4 -4
- package/dist/assets/webgpuProbe.js +127 -70
- package/dist/client/SignalingClient.d.ts.map +1 -1
- package/dist/client/light-preloader.d.ts +14 -1
- package/dist/client/light-preloader.d.ts.map +1 -1
- package/dist/consent/notice.d.ts +3 -1
- package/dist/consent/notice.d.ts.map +1 -1
- package/dist/executor/WorkerPool.d.ts +7 -0
- package/dist/executor/WorkerPool.d.ts.map +1 -1
- package/dist/executor/__tests__/resultLimits.test.d.ts +2 -0
- package/dist/executor/__tests__/resultLimits.test.d.ts.map +1 -0
- package/dist/executor/container-executor.d.ts +15 -0
- package/dist/executor/container-executor.d.ts.map +1 -1
- package/dist/executor/memoryProbe.d.ts +25 -1
- package/dist/executor/memoryProbe.d.ts.map +1 -1
- package/dist/executor/resultLimits.d.ts +19 -0
- package/dist/executor/resultLimits.d.ts.map +1 -0
- package/dist/index.js +357 -326
- package/dist/vector-store/HNSWIndex.d.ts +4 -1
- package/dist/vector-store/HNSWIndex.d.ts.map +1 -1
- package/dist/vector-store/VectorStoreEngine.d.ts +7 -1
- package/dist/vector-store/VectorStoreEngine.d.ts.map +1 -1
- package/dist/vector-store/__tests__/sharedEngine.test.d.ts +2 -0
- package/dist/vector-store/__tests__/sharedEngine.test.d.ts.map +1 -0
- package/dist/vector-store/sharedEngine.d.ts +13 -0
- package/dist/vector-store/sharedEngine.d.ts.map +1 -0
- package/dist/worker.js +187 -175
- package/dist/workloads/__tests__/cc-extract.test.d.ts +2 -0
- package/dist/workloads/__tests__/cc-extract.test.d.ts.map +1 -0
- package/dist/workloads/__tests__/decision.test.d.ts +2 -0
- package/dist/workloads/__tests__/decision.test.d.ts.map +1 -0
- package/dist/workloads/__tests__/embed-gemma.test.d.ts +2 -0
- package/dist/workloads/__tests__/embed-gemma.test.d.ts.map +1 -0
- package/dist/workloads/__tests__/nudenet-session-cache.test.d.ts +2 -0
- package/dist/workloads/__tests__/nudenet-session-cache.test.d.ts.map +1 -0
- package/dist/workloads/__tests__/wasm-threads.test.d.ts.map +1 -1
- package/dist/workloads/ai-inference.d.ts +21 -2
- package/dist/workloads/ai-inference.d.ts.map +1 -1
- package/dist/workloads/ai-model-registry.d.ts.map +1 -1
- package/dist/workloads/cc-extract.d.ts +11 -0
- package/dist/workloads/cc-extract.d.ts.map +1 -0
- package/dist/workloads/decision.d.ts +23 -0
- package/dist/workloads/decision.d.ts.map +1 -0
- package/dist/workloads/embed-gemma.d.ts +8 -0
- package/dist/workloads/embed-gemma.d.ts.map +1 -0
- package/dist/workloads/image-process.d.ts +8 -0
- package/dist/workloads/image-process.d.ts.map +1 -1
- package/dist/workloads/nudenet.d.ts.map +1 -1
- package/dist/workloads/vector-embed.d.ts +16 -0
- package/dist/workloads/vector-embed.d.ts.map +1 -1
- package/dist/workloads/vector-query.d.ts.map +1 -1
- package/dist/workloads/vector-store.d.ts.map +1 -1
- package/package.json +4 -3
package/README.md
CHANGED
|
@@ -170,6 +170,8 @@ initFlaxiaNode({ /* ... */ })
|
|
|
170
170
|
| `image-process` | `workloads/image-process.ts` | OffscreenCanvas API |
|
|
171
171
|
| `container` | `workloads/container.ts` | container2wasm + WASI |
|
|
172
172
|
| `nudenet` | `workloads/nudenet.ts` | NSFW(成人向け)画像判定 |
|
|
173
|
+
| `cc-extract` | `workloads/cc-extract.ts` | Common Crawl WARC の range 取得 + parse5 本文抽出 |
|
|
174
|
+
| `embed-gemma` | `workloads/embed-gemma.ts` | EmbeddingGemma 2(text-only, MRL) |
|
|
173
175
|
|
|
174
176
|
## フロー
|
|
175
177
|
|
|
@@ -1,122 +1,8 @@
|
|
|
1
1
|
import { t as e } from "./chunk.js";
|
|
2
|
-
import {
|
|
2
|
+
import { B as t, S as n } from "./webgpuProbe.js";
|
|
3
3
|
import { n as r, r as i, t as a } from "./transformers.web.js";
|
|
4
|
-
|
|
5
|
-
var
|
|
6
|
-
"Xenova/bert-base-multilingual-uncased-sentiment": {
|
|
7
|
-
repo: "Xenova/bert-base-multilingual-uncased-sentiment",
|
|
8
|
-
revision: "main",
|
|
9
|
-
maxBytes: o
|
|
10
|
-
},
|
|
11
|
-
"onnx-community/Qwen3-Embedding-0.6B-ONNX": {
|
|
12
|
-
repo: "onnx-community/Qwen3-Embedding-0.6B-ONNX",
|
|
13
|
-
revision: "main",
|
|
14
|
-
maxBytes: o
|
|
15
|
-
},
|
|
16
|
-
"Xenova/m2m100_418M": {
|
|
17
|
-
repo: "Xenova/m2m100_418M",
|
|
18
|
-
revision: "main",
|
|
19
|
-
maxBytes: o
|
|
20
|
-
},
|
|
21
|
-
"TeamFlaxia/plamo-3-translate-0.8b-onnx": {
|
|
22
|
-
repo: "TeamFlaxia/plamo-3-translate-0.8b-onnx",
|
|
23
|
-
revision: "1c8417ee1a72442440ad515b02f99186c81fd08a",
|
|
24
|
-
maxBytes: 2 * 1024 ** 3
|
|
25
|
-
}
|
|
26
|
-
}).map(([e, t]) => [e, { ...t }])), l = [
|
|
27
|
-
"wasm",
|
|
28
|
-
"cpu",
|
|
29
|
-
"webgpu"
|
|
30
|
-
], u = [
|
|
31
|
-
"q4",
|
|
32
|
-
"q4f16",
|
|
33
|
-
"q8",
|
|
34
|
-
"int8",
|
|
35
|
-
"uint8"
|
|
36
|
-
], d = {
|
|
37
|
-
max_new_tokens: {
|
|
38
|
-
min: 1,
|
|
39
|
-
max: 1024
|
|
40
|
-
},
|
|
41
|
-
temperature: {
|
|
42
|
-
min: 0,
|
|
43
|
-
max: 4
|
|
44
|
-
},
|
|
45
|
-
top_p: {
|
|
46
|
-
min: 0,
|
|
47
|
-
max: 1
|
|
48
|
-
},
|
|
49
|
-
top_k: {
|
|
50
|
-
min: 0,
|
|
51
|
-
max: 1e3
|
|
52
|
-
},
|
|
53
|
-
repetition_penalty: {
|
|
54
|
-
min: 0,
|
|
55
|
-
max: 4
|
|
56
|
-
}
|
|
57
|
-
}, f = 1e5;
|
|
58
|
-
function p() {
|
|
59
|
-
return [...c.keys()];
|
|
60
|
-
}
|
|
61
|
-
function m(e) {
|
|
62
|
-
if (typeof e != "string" || !e) throw Error(`ai-inference: model must be a non-empty string, got ${JSON.stringify(e)}`);
|
|
63
|
-
let t = c.get(e);
|
|
64
|
-
if (!t) throw Error(`ai-inference: model not allowed: ${e} (allowed: ${p().join(", ") || "none"})`);
|
|
65
|
-
return t;
|
|
66
|
-
}
|
|
67
|
-
function h(e) {
|
|
68
|
-
if (e != null) {
|
|
69
|
-
if (typeof e != "string" || !l.includes(e)) throw Error(`ai-inference: device not allowed: ${JSON.stringify(e)} (allowed: ${l.join(", ")})`);
|
|
70
|
-
return e;
|
|
71
|
-
}
|
|
72
|
-
}
|
|
73
|
-
function g(e) {
|
|
74
|
-
if (e != null) {
|
|
75
|
-
if (typeof e != "string" || !u.includes(e)) throw Error(`ai-inference: dtype not allowed: ${JSON.stringify(e)} (allowed: ${u.join(", ")})`);
|
|
76
|
-
return e;
|
|
77
|
-
}
|
|
78
|
-
}
|
|
79
|
-
function _(e) {
|
|
80
|
-
if (e == null) return !1;
|
|
81
|
-
if (typeof e != "boolean") throw Error(`ai-inference: option ondeviceOnly must be a boolean, got ${JSON.stringify(e)}`);
|
|
82
|
-
return e;
|
|
83
|
-
}
|
|
84
|
-
function v(e) {
|
|
85
|
-
if (typeof e == "string") {
|
|
86
|
-
if (e.length > 1e5) throw Error(`ai-inference: input exceeds ${f} characters`);
|
|
87
|
-
return e;
|
|
88
|
-
}
|
|
89
|
-
if (Array.isArray(e)) {
|
|
90
|
-
if (e.length === 0 || e.length > 64) throw Error("ai-inference: input array must hold 1..64 items");
|
|
91
|
-
for (let t of e) {
|
|
92
|
-
if (typeof t != "string") throw Error("ai-inference: input array items must be strings");
|
|
93
|
-
if (t.length > 1e5) throw Error(`ai-inference: input exceeds ${f} characters`);
|
|
94
|
-
}
|
|
95
|
-
return e;
|
|
96
|
-
}
|
|
97
|
-
throw Error("ai-inference: input must be a string or an array of strings");
|
|
98
|
-
}
|
|
99
|
-
function y(e, t) {
|
|
100
|
-
for (let [n, r] of Object.entries(d)) {
|
|
101
|
-
let i = t[n];
|
|
102
|
-
if (i != null) {
|
|
103
|
-
if (typeof i != "number" || !Number.isFinite(i) || i < r.min || i > r.max) throw Error(`ai-inference: option ${n} must be a number within ${r.min}..${r.max}, got ${JSON.stringify(i)}`);
|
|
104
|
-
e[n] = i;
|
|
105
|
-
}
|
|
106
|
-
}
|
|
107
|
-
if (t.do_sample !== void 0) {
|
|
108
|
-
if (typeof t.do_sample != "boolean") throw Error("ai-inference: option do_sample must be a boolean");
|
|
109
|
-
e.do_sample = t.do_sample;
|
|
110
|
-
}
|
|
111
|
-
for (let n of ["src_lang", "tgt_lang"]) {
|
|
112
|
-
let r = t[n];
|
|
113
|
-
if (r != null) {
|
|
114
|
-
if (typeof r != "string" || !/^[a-z]{2,3}(_[A-Za-z0-9]{2,8})?$/.test(r)) throw Error(`ai-inference: option ${n} must be a language code like 'en_XX'`);
|
|
115
|
-
e[n] = r;
|
|
116
|
-
}
|
|
117
|
-
}
|
|
118
|
-
}
|
|
119
|
-
var b = class {
|
|
4
|
+
import { a as o, i as s, n as c, o as l, r as u, s as d, t as f } from "./ai-model-registry.js";
|
|
5
|
+
var p = class {
|
|
120
6
|
constructor(e = {}) {
|
|
121
7
|
this.options = e, this.entries = /* @__PURE__ */ new Map();
|
|
122
8
|
}
|
|
@@ -236,7 +122,7 @@ var b = class {
|
|
|
236
122
|
return Math.max(1, this.options.ttlMs ?? 18e5);
|
|
237
123
|
}
|
|
238
124
|
};
|
|
239
|
-
function
|
|
125
|
+
function m(e) {
|
|
240
126
|
return JSON.stringify([
|
|
241
127
|
e.task,
|
|
242
128
|
e.modelId,
|
|
@@ -248,24 +134,24 @@ function x(e) {
|
|
|
248
134
|
}
|
|
249
135
|
//#endregion
|
|
250
136
|
//#region src/workloads/ai-inference.ts
|
|
251
|
-
var
|
|
252
|
-
DEFAULT_AI_MODEL_MAX_BYTES: () =>
|
|
253
|
-
LIGHT_INFERENCE_PLAMO_TRANSLATE_DTYPE: () =>
|
|
254
|
-
LIGHT_INFERENCE_PLAMO_TRANSLATE_MODEL: () =>
|
|
255
|
-
LIGHT_INFERENCE_PLAMO_TRANSLATE_TASK: () =>
|
|
256
|
-
LIGHT_INFERENCE_TRANSLATE_DTYPE: () =>
|
|
257
|
-
LIGHT_INFERENCE_TRANSLATE_MODEL: () =>
|
|
258
|
-
LIGHT_INFERENCE_TRANSLATE_TASK: () =>
|
|
259
|
-
MAX_AI_INPUT_CHARS: () =>
|
|
137
|
+
var h = /* @__PURE__ */ e({
|
|
138
|
+
DEFAULT_AI_MODEL_MAX_BYTES: () => DEFAULT_AI_MODEL_MAX_BYTES,
|
|
139
|
+
LIGHT_INFERENCE_PLAMO_TRANSLATE_DTYPE: () => T,
|
|
140
|
+
LIGHT_INFERENCE_PLAMO_TRANSLATE_MODEL: () => C,
|
|
141
|
+
LIGHT_INFERENCE_PLAMO_TRANSLATE_TASK: () => w,
|
|
142
|
+
LIGHT_INFERENCE_TRANSLATE_DTYPE: () => S,
|
|
143
|
+
LIGHT_INFERENCE_TRANSLATE_MODEL: () => b,
|
|
144
|
+
LIGHT_INFERENCE_TRANSLATE_TASK: () => x,
|
|
145
|
+
MAX_AI_INPUT_CHARS: () => MAX_AI_INPUT_CHARS,
|
|
260
146
|
MAX_AI_INPUT_ITEMS: () => 64,
|
|
261
147
|
MAX_PENDING_TOKEN_CALLBACKS: () => 64,
|
|
262
|
-
MAX_PENDING_TOKEN_CHARS: () =>
|
|
263
|
-
getModelCacheSnapshot: () =>
|
|
264
|
-
handleAiInference: () =>
|
|
265
|
-
isAiModelPreloaded: () =>
|
|
266
|
-
preloadAiModel: () =>
|
|
267
|
-
releaseCache: () =>
|
|
268
|
-
}),
|
|
148
|
+
MAX_PENDING_TOKEN_CHARS: () => v,
|
|
149
|
+
getModelCacheSnapshot: () => D,
|
|
150
|
+
handleAiInference: () => A,
|
|
151
|
+
isAiModelPreloaded: () => y,
|
|
152
|
+
preloadAiModel: () => j,
|
|
153
|
+
releaseCache: () => E
|
|
154
|
+
}), g = [
|
|
269
155
|
"text-classification",
|
|
270
156
|
"token-classification",
|
|
271
157
|
"question-answering",
|
|
@@ -291,9 +177,12 @@ var S = /* @__PURE__ */ e({
|
|
|
291
177
|
"depth-estimation",
|
|
292
178
|
"feature-extraction",
|
|
293
179
|
"image-feature-extraction"
|
|
294
|
-
],
|
|
295
|
-
|
|
296
|
-
|
|
180
|
+
], _ = new p(), v = 256 * 1024, y = (e) => {
|
|
181
|
+
let t = D();
|
|
182
|
+
return typeof e == "string" ? t.entries.some((t) => t.modelId === e && t.state === "ready") : t.entries.some((t) => t.modelId === e.modelId && t.revision === e.revision && t.dtype === e.dtype && t.backend === e.backend && t.state === "ready");
|
|
183
|
+
}, b = "Xenova/m2m100_418M", x = "translation", S = "q4f16", C = "TeamFlaxia/plamo-3-translate-0.8b-onnx", w = "text-generation", T = "q4f16", E = () => {
|
|
184
|
+
_.releaseAll();
|
|
185
|
+
}, D = () => _.getSnapshot(), O = (e, t) => {
|
|
297
186
|
let n = "", r = null, i = Promise.resolve(), a = 0, o = 0, s, c = !1, l = !1, u = (t) => {
|
|
298
187
|
if (a >= 64 || o + t.length > 262144) throw c = !0, s = /* @__PURE__ */ Error("ai-inference token consumer is too slow; generation stopped at the bounded stream buffer"), s;
|
|
299
188
|
a++, o += t.length, i = i.then(() => {
|
|
@@ -332,95 +221,106 @@ var S = /* @__PURE__ */ e({
|
|
|
332
221
|
cancel: () => p(!1)
|
|
333
222
|
};
|
|
334
223
|
};
|
|
335
|
-
function
|
|
224
|
+
function k() {
|
|
336
225
|
let e = globalThis.fetch.bind(globalThis);
|
|
337
226
|
r.fetch = async (t, n) => {
|
|
338
227
|
let r = new URL(String(t));
|
|
339
228
|
if (!/(^|\.)huggingface\.co$/.test(r.hostname) && !/(^|\.)hf\.co$/.test(r.hostname)) throw Error(`ai-inference: model download blocked from ${r.hostname}`);
|
|
340
229
|
let i = await e(r, n), a = Number(i.headers.get("content-length"));
|
|
341
|
-
if (Number.isFinite(a) && a >
|
|
230
|
+
if (Number.isFinite(a) && a > f) throw Error(`ai-inference: model file too large (${a} bytes > ${f})`);
|
|
342
231
|
return i;
|
|
343
232
|
};
|
|
344
233
|
}
|
|
345
|
-
|
|
346
|
-
var
|
|
347
|
-
let { task: r, model:
|
|
348
|
-
if (
|
|
349
|
-
|
|
350
|
-
|
|
351
|
-
|
|
234
|
+
k(), r.backends?.onnx?.wasm && (r.backends.onnx.wasm.numThreads = 1, r.backends.onnx.wasm.proxy = !1);
|
|
235
|
+
var A = async (e, n) => {
|
|
236
|
+
let { task: r, model: f, input: p, options: h = {} } = e;
|
|
237
|
+
if (e.decision) {
|
|
238
|
+
let { handleDecision: t } = await import("./decision.js");
|
|
239
|
+
return await t(e);
|
|
240
|
+
}
|
|
241
|
+
if (!g.includes(r)) throw Error(`Invalid or unsupported task: ${r}. Supported tasks are: ${g.join(", ")}`);
|
|
242
|
+
let v = l(f), b = s(h.device), x = o(h.dtype), S = d(h.ondeviceOnly), C = u(p), w = {};
|
|
243
|
+
c(w, h);
|
|
244
|
+
let T = x ?? "q4f16", E = b ?? "wasm";
|
|
245
|
+
if (S && !y({
|
|
246
|
+
modelId: f,
|
|
247
|
+
revision: v.revision,
|
|
248
|
+
dtype: T,
|
|
249
|
+
backend: E
|
|
250
|
+
})) throw new t(`ondeviceOnly: model ${f} is not preloaded on this device`, "preloading");
|
|
251
|
+
let D = m({
|
|
352
252
|
task: r,
|
|
353
|
-
modelId:
|
|
354
|
-
revision:
|
|
355
|
-
dtype:
|
|
356
|
-
backend:
|
|
357
|
-
}),
|
|
358
|
-
modelId:
|
|
359
|
-
revision:
|
|
360
|
-
dtype:
|
|
361
|
-
backend:
|
|
253
|
+
modelId: f,
|
|
254
|
+
revision: v.revision,
|
|
255
|
+
dtype: T,
|
|
256
|
+
backend: E
|
|
257
|
+
}), k = performance.now(), A = await _.acquire(D, {
|
|
258
|
+
modelId: f,
|
|
259
|
+
revision: v.revision,
|
|
260
|
+
dtype: T,
|
|
261
|
+
backend: E
|
|
362
262
|
}, async () => {
|
|
363
263
|
let e = {
|
|
364
|
-
dtype:
|
|
365
|
-
device:
|
|
366
|
-
revision:
|
|
264
|
+
dtype: T,
|
|
265
|
+
device: b,
|
|
266
|
+
revision: v.revision
|
|
367
267
|
};
|
|
368
|
-
console.log(`[flaxia-node] ai-inference: loading pipeline task=${r} model=${
|
|
268
|
+
console.log(`[flaxia-node] ai-inference: loading pipeline task=${r} model=${v.repo} revision=${v.revision} dtype=${e.dtype} device=${b ?? "wasm"}`);
|
|
369
269
|
let t;
|
|
370
270
|
try {
|
|
371
|
-
t = await i(r,
|
|
271
|
+
t = await i(r, v.repo, e);
|
|
372
272
|
} catch (n) {
|
|
373
|
-
if (
|
|
374
|
-
else throw console.error(`[flaxia-node] ai-inference: pipeline load FAILED task=${r} model=${
|
|
273
|
+
if (b && b !== "wasm") console.warn(`[AiInference] ${b} failed, falling back to wasm:`, n), e.device = "wasm", t = await i(r, v.repo, e);
|
|
274
|
+
else throw console.error(`[flaxia-node] ai-inference: pipeline load FAILED task=${r} model=${v.repo} error=${n instanceof Error ? n.message : String(n)}`), n;
|
|
375
275
|
}
|
|
376
|
-
return console.log(`[flaxia-node] ai-inference: pipeline ready task=${r} model=${
|
|
276
|
+
return console.log(`[flaxia-node] ai-inference: pipeline ready task=${r} model=${v.repo} loadMs=${Math.round(performance.now() - k)}`), {
|
|
377
277
|
value: t,
|
|
378
278
|
descriptor: { backend: e.device ?? "wasm" }
|
|
379
279
|
};
|
|
380
|
-
}, (e) => e.dispose?.()),
|
|
280
|
+
}, (e) => e.dispose?.()), j = A.value;
|
|
381
281
|
try {
|
|
382
282
|
let e = {
|
|
383
|
-
max_new_tokens:
|
|
384
|
-
do_sample:
|
|
283
|
+
max_new_tokens: w.max_new_tokens ?? 128,
|
|
284
|
+
do_sample: w.do_sample ?? !1
|
|
385
285
|
};
|
|
386
|
-
for (let [t, n] of Object.entries(
|
|
387
|
-
let t = n &&
|
|
388
|
-
t && (e.streamer = new a(
|
|
286
|
+
for (let [t, n] of Object.entries(w)) t === "max_new_tokens" || t === "do_sample" || (e[t] = n);
|
|
287
|
+
let t = n && j.tokenizer ? O(n, h) : null;
|
|
288
|
+
t && (e.streamer = new a(j.tokenizer, {
|
|
389
289
|
skip_prompt: !0,
|
|
390
290
|
skip_special_tokens: !0,
|
|
391
291
|
callback_function: t.callback
|
|
392
292
|
}));
|
|
393
293
|
let i = performance.now(), o;
|
|
394
294
|
try {
|
|
395
|
-
o = await
|
|
295
|
+
o = await j(C, e);
|
|
396
296
|
} catch (e) {
|
|
397
|
-
throw /out of memory|oom|allocation failed|could not allocate/i.test(e instanceof Error ? e.message : String(e)) &&
|
|
297
|
+
throw /out of memory|oom|allocation failed|could not allocate/i.test(e instanceof Error ? e.message : String(e)) && A.invalidate(), t && await t.cancel(), e;
|
|
398
298
|
}
|
|
399
|
-
return await t?.finish(), console.log(`[flaxia-node] ai-inference: done task=${r} model=${
|
|
299
|
+
return await t?.finish(), console.log(`[flaxia-node] ai-inference: done task=${r} model=${v.repo} execMs=${Math.round(performance.now() - i)} inputLen=${String(C).length}`), { output: o };
|
|
400
300
|
} finally {
|
|
401
|
-
await
|
|
301
|
+
await A.release();
|
|
402
302
|
}
|
|
403
|
-
},
|
|
404
|
-
let { task: t, model: n } = e, r =
|
|
303
|
+
}, j = async (e) => {
|
|
304
|
+
let { task: t, model: n } = e, r = l(n), a = s(e.device), c = o(e.dtype) ?? "q4f16", u = a ?? "wasm", d = m({
|
|
405
305
|
task: t,
|
|
406
306
|
modelId: n,
|
|
407
307
|
revision: r.revision,
|
|
408
|
-
dtype:
|
|
409
|
-
backend:
|
|
308
|
+
dtype: c,
|
|
309
|
+
backend: u
|
|
410
310
|
});
|
|
411
|
-
await (await
|
|
311
|
+
await (await _.acquire(d, {
|
|
412
312
|
modelId: n,
|
|
413
313
|
revision: r.revision,
|
|
414
|
-
dtype:
|
|
415
|
-
backend:
|
|
314
|
+
dtype: c,
|
|
315
|
+
backend: u
|
|
416
316
|
}, async () => (console.log(`[flaxia-node] ai-inference: preloading task=${t} model=${r.repo}`), {
|
|
417
317
|
value: await i(t, r.repo, {
|
|
418
|
-
dtype:
|
|
318
|
+
dtype: c,
|
|
419
319
|
device: a,
|
|
420
320
|
revision: r.revision
|
|
421
321
|
}),
|
|
422
|
-
descriptor: { backend:
|
|
322
|
+
descriptor: { backend: u }
|
|
423
323
|
}), (e) => e.dispose?.())).release();
|
|
424
324
|
};
|
|
425
325
|
//#endregion
|
|
426
|
-
export {
|
|
326
|
+
export { b as a, y as c, S as i, j as l, C as n, x as o, w as r, h as s, T as t };
|
|
@@ -0,0 +1,127 @@
|
|
|
1
|
+
//#region src/workloads/ai-model-registry.ts
|
|
2
|
+
var e = 512 * 1024 * 1024, t = 4 * 1024 ** 3, n = new Map(Object.entries({
|
|
3
|
+
"Xenova/bert-base-multilingual-uncased-sentiment": {
|
|
4
|
+
repo: "Xenova/bert-base-multilingual-uncased-sentiment",
|
|
5
|
+
revision: "main",
|
|
6
|
+
maxBytes: e
|
|
7
|
+
},
|
|
8
|
+
"onnx-community/Qwen3-Embedding-0.6B-ONNX": {
|
|
9
|
+
repo: "onnx-community/Qwen3-Embedding-0.6B-ONNX",
|
|
10
|
+
revision: "main",
|
|
11
|
+
maxBytes: e
|
|
12
|
+
},
|
|
13
|
+
"LiquidAI/d1-omni-600M": {
|
|
14
|
+
repo: "LiquidAI/d1-omni-600M",
|
|
15
|
+
revision: "02b55d7076f15129e59ab3f94783f32c4b088674",
|
|
16
|
+
maxBytes: 3 * 1024 ** 3
|
|
17
|
+
},
|
|
18
|
+
"onnx-community/embeddinggemma-2-ONNX": {
|
|
19
|
+
repo: "onnx-community/embeddinggemma-2-ONNX",
|
|
20
|
+
revision: "main",
|
|
21
|
+
maxBytes: 1024 * 1024 * 1024
|
|
22
|
+
},
|
|
23
|
+
"Xenova/m2m100_418M": {
|
|
24
|
+
repo: "Xenova/m2m100_418M",
|
|
25
|
+
revision: "main",
|
|
26
|
+
maxBytes: e
|
|
27
|
+
},
|
|
28
|
+
"TeamFlaxia/plamo-3-translate-0.8b-onnx": {
|
|
29
|
+
repo: "TeamFlaxia/plamo-3-translate-0.8b-onnx",
|
|
30
|
+
revision: "1c8417ee1a72442440ad515b02f99186c81fd08a",
|
|
31
|
+
maxBytes: 2 * 1024 ** 3
|
|
32
|
+
}
|
|
33
|
+
}).map(([e, t]) => [e, { ...t }])), r = [
|
|
34
|
+
"wasm",
|
|
35
|
+
"cpu",
|
|
36
|
+
"webgpu"
|
|
37
|
+
], i = [
|
|
38
|
+
"q4",
|
|
39
|
+
"q4f16",
|
|
40
|
+
"q8",
|
|
41
|
+
"int8",
|
|
42
|
+
"uint8"
|
|
43
|
+
], a = {
|
|
44
|
+
max_new_tokens: {
|
|
45
|
+
min: 1,
|
|
46
|
+
max: 1024
|
|
47
|
+
},
|
|
48
|
+
temperature: {
|
|
49
|
+
min: 0,
|
|
50
|
+
max: 4
|
|
51
|
+
},
|
|
52
|
+
top_p: {
|
|
53
|
+
min: 0,
|
|
54
|
+
max: 1
|
|
55
|
+
},
|
|
56
|
+
top_k: {
|
|
57
|
+
min: 0,
|
|
58
|
+
max: 1e3
|
|
59
|
+
},
|
|
60
|
+
repetition_penalty: {
|
|
61
|
+
min: 0,
|
|
62
|
+
max: 4
|
|
63
|
+
}
|
|
64
|
+
}, o = 1e5;
|
|
65
|
+
function s() {
|
|
66
|
+
return [...n.keys()];
|
|
67
|
+
}
|
|
68
|
+
function c(e) {
|
|
69
|
+
if (typeof e != "string" || !e) throw Error(`ai-inference: model must be a non-empty string, got ${JSON.stringify(e)}`);
|
|
70
|
+
let t = n.get(e);
|
|
71
|
+
if (!t) throw Error(`ai-inference: model not allowed: ${e} (allowed: ${s().join(", ") || "none"})`);
|
|
72
|
+
return t;
|
|
73
|
+
}
|
|
74
|
+
function l(e) {
|
|
75
|
+
if (e != null) {
|
|
76
|
+
if (typeof e != "string" || !r.includes(e)) throw Error(`ai-inference: device not allowed: ${JSON.stringify(e)} (allowed: ${r.join(", ")})`);
|
|
77
|
+
return e;
|
|
78
|
+
}
|
|
79
|
+
}
|
|
80
|
+
function u(e) {
|
|
81
|
+
if (e != null) {
|
|
82
|
+
if (typeof e != "string" || !i.includes(e)) throw Error(`ai-inference: dtype not allowed: ${JSON.stringify(e)} (allowed: ${i.join(", ")})`);
|
|
83
|
+
return e;
|
|
84
|
+
}
|
|
85
|
+
}
|
|
86
|
+
function d(e) {
|
|
87
|
+
if (e == null) return !1;
|
|
88
|
+
if (typeof e != "boolean") throw Error(`ai-inference: option ondeviceOnly must be a boolean, got ${JSON.stringify(e)}`);
|
|
89
|
+
return e;
|
|
90
|
+
}
|
|
91
|
+
function f(e) {
|
|
92
|
+
if (typeof e == "string") {
|
|
93
|
+
if (e.length > 1e5) throw Error(`ai-inference: input exceeds ${o} characters`);
|
|
94
|
+
return e;
|
|
95
|
+
}
|
|
96
|
+
if (Array.isArray(e)) {
|
|
97
|
+
if (e.length === 0 || e.length > 64) throw Error("ai-inference: input array must hold 1..64 items");
|
|
98
|
+
for (let t of e) {
|
|
99
|
+
if (typeof t != "string") throw Error("ai-inference: input array items must be strings");
|
|
100
|
+
if (t.length > 1e5) throw Error(`ai-inference: input exceeds ${o} characters`);
|
|
101
|
+
}
|
|
102
|
+
return e;
|
|
103
|
+
}
|
|
104
|
+
throw Error("ai-inference: input must be a string or an array of strings");
|
|
105
|
+
}
|
|
106
|
+
function p(e, t) {
|
|
107
|
+
for (let [n, r] of Object.entries(a)) {
|
|
108
|
+
let i = t[n];
|
|
109
|
+
if (i != null) {
|
|
110
|
+
if (typeof i != "number" || !Number.isFinite(i) || i < r.min || i > r.max) throw Error(`ai-inference: option ${n} must be a number within ${r.min}..${r.max}, got ${JSON.stringify(i)}`);
|
|
111
|
+
e[n] = i;
|
|
112
|
+
}
|
|
113
|
+
}
|
|
114
|
+
if (t.do_sample !== void 0) {
|
|
115
|
+
if (typeof t.do_sample != "boolean") throw Error("ai-inference: option do_sample must be a boolean");
|
|
116
|
+
e.do_sample = t.do_sample;
|
|
117
|
+
}
|
|
118
|
+
for (let n of ["src_lang", "tgt_lang"]) {
|
|
119
|
+
let r = t[n];
|
|
120
|
+
if (r != null) {
|
|
121
|
+
if (typeof r != "string" || !/^[a-z]{2,3}(_[A-Za-z0-9]{2,8})?$/.test(r)) throw Error(`ai-inference: option ${n} must be a language code like 'en_XX'`);
|
|
122
|
+
e[n] = r;
|
|
123
|
+
}
|
|
124
|
+
}
|
|
125
|
+
}
|
|
126
|
+
//#endregion
|
|
127
|
+
export { u as a, l as i, p as n, c as o, f as r, d as s, t };
|