Repository navigation
Expand file tree
/
Copy pathinferenceWorker.ts
More file actions
191 lines (169 loc) · 5.34 KB
/
Copy pathinferenceWorker.ts
File metadata and controls
191 lines (169 loc) · 5.34 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
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
import * as ort from "onnxruntime-web/wasm";
import type { RuntimeBackend } from "./modelProfiles";
ort.env.wasm.numThreads = 1;
type EmbedMessage = {
type: "embed";
id: string;
modelUrl: string;
inputSize: number;
input: Float32Array;
backendCandidates: RuntimeBackend[];
};
type ModelBytes = {
bytes: Uint8Array;
cacheHit: boolean;
modelLoadMs: number;
};
type SessionLoad = {
backend: RuntimeBackend;
session: ort.InferenceSession;
cacheHit: boolean;
modelLoadMs: number;
sessionCreateMs: number;
sessionReused: boolean;
};
type AttemptFailure = {
backend: RuntimeBackend;
error: string;
};
const MODEL_CACHE_NAME = "path-finder-models-v1";
const modelBytePromises = new Map<string, Promise<ModelBytes>>();
const sessionPromises = new Map<string, Promise<SessionLoad>>();
async function loadModelBytes(modelUrl: string): Promise<ModelBytes> {
const existing = modelBytePromises.get(modelUrl);
if (existing) {
const cached = await existing;
return { ...cached, modelLoadMs: 0 };
}
const promise = (async () => {
const startedAt = performance.now();
const request = new Request(modelUrl, { cache: "force-cache" });
const modelCache = "caches" in self ? await caches.open(MODEL_CACHE_NAME) : null;
const cachedResponse = await modelCache?.match(request);
if (cachedResponse) {
return {
bytes: new Uint8Array(await cachedResponse.arrayBuffer()),
cacheHit: true,
modelLoadMs: performance.now() - startedAt,
};
}
const response = await fetch(request);
if (!response.ok) {
throw new Error(`모델 다운로드 실패: ${response.status}`);
}
await modelCache?.put(request, response.clone());
return {
bytes: new Uint8Array(await response.arrayBuffer()),
cacheHit: false,
modelLoadMs: performance.now() - startedAt,
};
})();
modelBytePromises.set(modelUrl, promise);
return promise;
}
async function createSession(modelUrl: string, backend: RuntimeBackend): Promise<SessionLoad> {
if (backend !== "wasm") {
throw new Error(`${backend}는 WASM Worker에서 실행할 수 없습니다.`);
}
const model = await loadModelBytes(modelUrl);
const startedAt = performance.now();
const session = await ort.InferenceSession.create(model.bytes, {
executionProviders: [backend],
graphOptimizationLevel: "all",
});
return {
backend,
session,
cacheHit: model.cacheHit,
modelLoadMs: model.modelLoadMs,
sessionCreateMs: performance.now() - startedAt,
sessionReused: false,
};
}
async function loadSession(modelUrl: string, backend: RuntimeBackend): Promise<SessionLoad> {
const key = `${backend}:${modelUrl}`;
const existing = sessionPromises.get(key);
if (existing) {
const loaded = await existing;
return {
...loaded,
modelLoadMs: 0,
sessionCreateMs: 0,
sessionReused: true,
};
}
const promise = createSession(modelUrl, backend);
sessionPromises.set(key, promise);
try {
return await promise;
} catch (error) {
sessionPromises.delete(key);
throw error;
}
}
async function loadFirstAvailableSession(
modelUrl: string,
backendCandidates: RuntimeBackend[],
): Promise<SessionLoad & { failures: AttemptFailure[] }> {
const failures: AttemptFailure[] = [];
for (const backend of backendCandidates) {
try {
return { ...(await loadSession(modelUrl, backend)), failures };
} catch (error) {
failures.push({
backend,
error: error instanceof Error ? error.message : "세션 생성 실패",
});
}
}
throw new Error(failures.map((failure) => `${failure.backend}: ${failure.error}`).join(" / "));
}
self.addEventListener("message", (event: MessageEvent<EmbedMessage>) => {
const message = event.data;
if (message.type !== "embed") {
return;
}
void (async () => {
try {
const sessionLoad = await loadFirstAvailableSession(
message.modelUrl,
message.backendCandidates.length > 0 ? message.backendCandidates : ["wasm"],
);
const tensor = new ort.Tensor("float32", message.input, [
1,
3,
message.inputSize,
message.inputSize,
]);
const inputName = sessionLoad.session.inputNames[0];
const outputName = sessionLoad.session.outputNames[0];
const inferenceStartedAt = performance.now();
const outputs = await sessionLoad.session.run({ [inputName]: tensor });
const inferenceMs = performance.now() - inferenceStartedAt;
const embedding = outputs.embedding ?? outputs[outputName];
if (!embedding || !(embedding.data instanceof Float32Array)) {
throw new Error("ONNX 모델 출력이 올바르지 않습니다.");
}
self.postMessage({
type: "embedding",
id: message.id,
values: Array.from(embedding.data),
backend: sessionLoad.backend,
timings: {
modelLoadMs: sessionLoad.modelLoadMs,
sessionCreateMs: sessionLoad.sessionCreateMs,
inferenceMs,
},
cacheHit: sessionLoad.cacheHit,
sessionReused: sessionLoad.sessionReused,
fallbackErrors: sessionLoad.failures,
});
} catch (error) {
self.postMessage({
type: "error",
id: message.id,
error: error instanceof Error ? error.message : "추론 중 오류가 발생했습니다.",
});
}
})();
});