@tuned-tensor/cli 0.4.14 → 0.4.15
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/dist/index.js +593 -17
- package/dist/index.js.map +1 -1
- package/package.json +1 -1
package/dist/index.js
CHANGED
|
@@ -595,6 +595,13 @@ function formatEvalScore(run) {
|
|
|
595
595
|
const score = getEvalScore(run);
|
|
596
596
|
return score == null ? void 0 : (score * 100).toFixed(1) + "%";
|
|
597
597
|
}
|
|
598
|
+
function formatPercent(rate) {
|
|
599
|
+
return rate == null ? void 0 : (rate * 100).toFixed(1) + "%";
|
|
600
|
+
}
|
|
601
|
+
function formatCountRate(count, total, rate) {
|
|
602
|
+
if (count == null || total == null || rate == null) return void 0;
|
|
603
|
+
return `${formatPercent(rate)} (${count}/${total})`;
|
|
604
|
+
}
|
|
598
605
|
function formatMinutes(minutes) {
|
|
599
606
|
if (minutes == null) return void 0;
|
|
600
607
|
if (minutes < 60) return `${Math.max(1, Math.round(minutes))}m`;
|
|
@@ -632,6 +639,78 @@ function printDiagnostics(diagnostics) {
|
|
|
632
639
|
console.log(` - ${insight}`);
|
|
633
640
|
}
|
|
634
641
|
}
|
|
642
|
+
printEvalOutputDiagnostics(diagnostics.output_diagnostics);
|
|
643
|
+
}
|
|
644
|
+
function printEvalOutputDiagnostics(diagnostics) {
|
|
645
|
+
if (!diagnostics) return;
|
|
646
|
+
const candidate = diagnostics.candidate;
|
|
647
|
+
const baseline = diagnostics.baseline;
|
|
648
|
+
console.log("\nOutput Diagnostics:");
|
|
649
|
+
printDetail([
|
|
650
|
+
["Tuned Valid JSON", formatCountRate(candidate.valid_json_count, candidate.total, candidate.valid_json_rate)],
|
|
651
|
+
["Tuned Strict JSON", formatCountRate(candidate.strict_json_count, candidate.total, candidate.strict_json_rate)],
|
|
652
|
+
[
|
|
653
|
+
"Tuned Schema Keys",
|
|
654
|
+
formatCountRate(candidate.expected_schema_keys_count, candidate.total, candidate.expected_schema_keys_rate)
|
|
655
|
+
],
|
|
656
|
+
[
|
|
657
|
+
"Tuned Non-JSON Prefix",
|
|
658
|
+
formatCountRate(candidate.non_json_prefix_count, candidate.total, candidate.non_json_prefix_rate)
|
|
659
|
+
],
|
|
660
|
+
[
|
|
661
|
+
"Tuned Reasoning Prefix",
|
|
662
|
+
formatCountRate(
|
|
663
|
+
candidate.visible_reasoning_prefix_count,
|
|
664
|
+
candidate.total,
|
|
665
|
+
candidate.visible_reasoning_prefix_rate
|
|
666
|
+
)
|
|
667
|
+
],
|
|
668
|
+
["Tuned Avg Output", candidate.avg_output_chars == null ? void 0 : `${candidate.avg_output_chars} chars`],
|
|
669
|
+
["Base Valid JSON", formatCountRate(baseline.valid_json_count, baseline.total, baseline.valid_json_rate)],
|
|
670
|
+
[
|
|
671
|
+
"Base Reasoning Prefix",
|
|
672
|
+
formatCountRate(
|
|
673
|
+
baseline.visible_reasoning_prefix_count,
|
|
674
|
+
baseline.total,
|
|
675
|
+
baseline.visible_reasoning_prefix_rate
|
|
676
|
+
)
|
|
677
|
+
]
|
|
678
|
+
]);
|
|
679
|
+
if (diagnostics.test) {
|
|
680
|
+
console.log("\nTest Output Diagnostics:");
|
|
681
|
+
printDetail([
|
|
682
|
+
[
|
|
683
|
+
"Tuned Valid JSON",
|
|
684
|
+
formatCountRate(
|
|
685
|
+
diagnostics.test.candidate.valid_json_count,
|
|
686
|
+
diagnostics.test.candidate.total,
|
|
687
|
+
diagnostics.test.candidate.valid_json_rate
|
|
688
|
+
)
|
|
689
|
+
],
|
|
690
|
+
[
|
|
691
|
+
"Tuned Strict JSON",
|
|
692
|
+
formatCountRate(
|
|
693
|
+
diagnostics.test.candidate.strict_json_count,
|
|
694
|
+
diagnostics.test.candidate.total,
|
|
695
|
+
diagnostics.test.candidate.strict_json_rate
|
|
696
|
+
)
|
|
697
|
+
],
|
|
698
|
+
[
|
|
699
|
+
"Tuned Reasoning Prefix",
|
|
700
|
+
formatCountRate(
|
|
701
|
+
diagnostics.test.candidate.visible_reasoning_prefix_count,
|
|
702
|
+
diagnostics.test.candidate.total,
|
|
703
|
+
diagnostics.test.candidate.visible_reasoning_prefix_rate
|
|
704
|
+
)
|
|
705
|
+
]
|
|
706
|
+
]);
|
|
707
|
+
}
|
|
708
|
+
if (diagnostics.insights?.length) {
|
|
709
|
+
console.log("\nEvaluation Insights:");
|
|
710
|
+
for (const insight of diagnostics.insights) {
|
|
711
|
+
console.log(` - ${insight}`);
|
|
712
|
+
}
|
|
713
|
+
}
|
|
635
714
|
}
|
|
636
715
|
function registerRunsCommands(parent) {
|
|
637
716
|
const runs = parent.command("runs").description("Manage runs");
|
|
@@ -687,6 +766,7 @@ function registerRunsCommands(parent) {
|
|
|
687
766
|
console.log(` ${k}: ${v}`);
|
|
688
767
|
}
|
|
689
768
|
}
|
|
769
|
+
printEvalOutputDiagnostics(data.eval_summary?.output_diagnostics);
|
|
690
770
|
if (data._events?.length) {
|
|
691
771
|
console.log("\nLatest Updates:");
|
|
692
772
|
for (const event of data._events.slice(-5)) {
|
|
@@ -1009,11 +1089,11 @@ function registerDatasetsCommands(parent) {
|
|
|
1009
1089
|
}
|
|
1010
1090
|
|
|
1011
1091
|
// src/commands/models.ts
|
|
1012
|
-
import { spawn, execFileSync } from "child_process";
|
|
1092
|
+
import { spawn as spawn2, execFileSync } from "child_process";
|
|
1013
1093
|
import { createHash } from "crypto";
|
|
1014
1094
|
import {
|
|
1015
1095
|
existsSync as existsSync4,
|
|
1016
|
-
mkdirSync as
|
|
1096
|
+
mkdirSync as mkdirSync3,
|
|
1017
1097
|
statSync as statSync3,
|
|
1018
1098
|
createWriteStream,
|
|
1019
1099
|
unlinkSync,
|
|
@@ -1023,7 +1103,7 @@ import {
|
|
|
1023
1103
|
rmSync
|
|
1024
1104
|
} from "fs";
|
|
1025
1105
|
import { homedir as homedir2 } from "os";
|
|
1026
|
-
import { dirname, join as join2, resolve as resolve2, basename as basename2 } from "path";
|
|
1106
|
+
import { dirname as dirname2, join as join2, resolve as resolve2, basename as basename2 } from "path";
|
|
1027
1107
|
import { Readable, Transform } from "stream";
|
|
1028
1108
|
import { pipeline } from "stream/promises";
|
|
1029
1109
|
|
|
@@ -1074,6 +1154,444 @@ Run \`tt init\` to create one.`
|
|
|
1074
1154
|
return raw;
|
|
1075
1155
|
}
|
|
1076
1156
|
|
|
1157
|
+
// src/serve-manager.ts
|
|
1158
|
+
import { spawn } from "child_process";
|
|
1159
|
+
import { randomUUID } from "crypto";
|
|
1160
|
+
import { appendFileSync, mkdirSync as mkdirSync2 } from "fs";
|
|
1161
|
+
import { createServer } from "http";
|
|
1162
|
+
import { createServer as createNetServer } from "net";
|
|
1163
|
+
import { dirname } from "path";
|
|
1164
|
+
var COMPLETION_ROUTES = /* @__PURE__ */ new Set(["/v1/chat/completions", "/chat/completions"]);
|
|
1165
|
+
function parsePositiveInteger(value, fallback) {
|
|
1166
|
+
return Number.isInteger(value) && value >= 0 ? value : fallback;
|
|
1167
|
+
}
|
|
1168
|
+
function sleep(ms) {
|
|
1169
|
+
return new Promise((resolve4) => setTimeout(resolve4, ms));
|
|
1170
|
+
}
|
|
1171
|
+
function readRequestBody(req) {
|
|
1172
|
+
return new Promise((resolve4, reject) => {
|
|
1173
|
+
const chunks = [];
|
|
1174
|
+
req.on("data", (chunk) => chunks.push(Buffer.from(chunk)));
|
|
1175
|
+
req.on("end", () => resolve4(Buffer.concat(chunks)));
|
|
1176
|
+
req.on("error", reject);
|
|
1177
|
+
});
|
|
1178
|
+
}
|
|
1179
|
+
function sendJson(res, status, payload) {
|
|
1180
|
+
const body = Buffer.from(JSON.stringify(payload), "utf8");
|
|
1181
|
+
res.writeHead(status, {
|
|
1182
|
+
"content-type": "application/json",
|
|
1183
|
+
"content-length": String(body.byteLength),
|
|
1184
|
+
"access-control-allow-origin": "*",
|
|
1185
|
+
"access-control-allow-headers": "authorization, content-type",
|
|
1186
|
+
"access-control-allow-methods": "GET, POST, OPTIONS"
|
|
1187
|
+
});
|
|
1188
|
+
res.end(body);
|
|
1189
|
+
}
|
|
1190
|
+
function copyResponseHeaders(headers) {
|
|
1191
|
+
const copied = {
|
|
1192
|
+
"access-control-allow-origin": "*",
|
|
1193
|
+
"access-control-allow-headers": "authorization, content-type",
|
|
1194
|
+
"access-control-allow-methods": "GET, POST, OPTIONS"
|
|
1195
|
+
};
|
|
1196
|
+
const contentType = headers.get("content-type");
|
|
1197
|
+
if (contentType) copied["content-type"] = contentType;
|
|
1198
|
+
return copied;
|
|
1199
|
+
}
|
|
1200
|
+
function tryJsonParse(value) {
|
|
1201
|
+
try {
|
|
1202
|
+
return JSON.parse(value);
|
|
1203
|
+
} catch {
|
|
1204
|
+
return null;
|
|
1205
|
+
}
|
|
1206
|
+
}
|
|
1207
|
+
function readPath(value, path) {
|
|
1208
|
+
if (!path) return null;
|
|
1209
|
+
let current = value;
|
|
1210
|
+
for (const part of path.split(".")) {
|
|
1211
|
+
if (!part) return null;
|
|
1212
|
+
if (!current || typeof current !== "object" || !(part in current)) return null;
|
|
1213
|
+
current = current[part];
|
|
1214
|
+
}
|
|
1215
|
+
return current ?? null;
|
|
1216
|
+
}
|
|
1217
|
+
function extractAssistantContent(response) {
|
|
1218
|
+
if (!response || typeof response !== "object") return null;
|
|
1219
|
+
const choices = response.choices;
|
|
1220
|
+
if (!Array.isArray(choices) || choices.length === 0) return null;
|
|
1221
|
+
const first = choices[0];
|
|
1222
|
+
if (!first || typeof first !== "object") return null;
|
|
1223
|
+
const message = first.message;
|
|
1224
|
+
if (!message || typeof message !== "object") return null;
|
|
1225
|
+
const content = message.content;
|
|
1226
|
+
return typeof content === "string" ? content : null;
|
|
1227
|
+
}
|
|
1228
|
+
function usageNumber(response, key) {
|
|
1229
|
+
if (!response || typeof response !== "object") return null;
|
|
1230
|
+
const usage = response.usage;
|
|
1231
|
+
if (!usage || typeof usage !== "object") return null;
|
|
1232
|
+
const value = usage[key];
|
|
1233
|
+
return typeof value === "number" && Number.isFinite(value) ? value : null;
|
|
1234
|
+
}
|
|
1235
|
+
function errorSummary(response, fallback) {
|
|
1236
|
+
if (!response || typeof response !== "object") return fallback;
|
|
1237
|
+
const error = response.error;
|
|
1238
|
+
if (!error || typeof error !== "object") return fallback;
|
|
1239
|
+
const message = error.message;
|
|
1240
|
+
return typeof message === "string" ? message : fallback;
|
|
1241
|
+
}
|
|
1242
|
+
async function allocatePort(host) {
|
|
1243
|
+
return new Promise((resolve4, reject) => {
|
|
1244
|
+
const server = createNetServer();
|
|
1245
|
+
server.on("error", reject);
|
|
1246
|
+
server.listen(0, host, () => {
|
|
1247
|
+
const address = server.address();
|
|
1248
|
+
if (!address || typeof address === "string") {
|
|
1249
|
+
server.close(() => reject(new Error("Could not allocate an internal port")));
|
|
1250
|
+
return;
|
|
1251
|
+
}
|
|
1252
|
+
const port = address.port;
|
|
1253
|
+
server.close(() => resolve4(port));
|
|
1254
|
+
});
|
|
1255
|
+
});
|
|
1256
|
+
}
|
|
1257
|
+
var ManagedModelServer = class {
|
|
1258
|
+
constructor(options) {
|
|
1259
|
+
this.options = options;
|
|
1260
|
+
this.spawnModelServer = options.spawnModelServer ?? spawn;
|
|
1261
|
+
this.idleTimeoutMs = parsePositiveInteger(options.idleTimeoutSeconds, 300) * 1e3;
|
|
1262
|
+
this.restartAfterRequests = parsePositiveInteger(options.restartAfterRequests, 100);
|
|
1263
|
+
this.logRecordOverride = options.logRecord;
|
|
1264
|
+
}
|
|
1265
|
+
spawnModelServer;
|
|
1266
|
+
idleTimeoutMs;
|
|
1267
|
+
restartAfterRequests;
|
|
1268
|
+
logRecordOverride;
|
|
1269
|
+
server = null;
|
|
1270
|
+
child = null;
|
|
1271
|
+
internalPort = null;
|
|
1272
|
+
queue = Promise.resolve();
|
|
1273
|
+
idleTimer = null;
|
|
1274
|
+
activeRequests = 0;
|
|
1275
|
+
queuedRequests = 0;
|
|
1276
|
+
requestsSinceStart = 0;
|
|
1277
|
+
restartCount = 0;
|
|
1278
|
+
stoppingChild = false;
|
|
1279
|
+
get isModelRunning() {
|
|
1280
|
+
return this.child != null;
|
|
1281
|
+
}
|
|
1282
|
+
get restarts() {
|
|
1283
|
+
return this.restartCount;
|
|
1284
|
+
}
|
|
1285
|
+
async start() {
|
|
1286
|
+
if (this.server) return;
|
|
1287
|
+
this.internalPort = await allocatePort("127.0.0.1");
|
|
1288
|
+
this.server = createServer((req, res) => {
|
|
1289
|
+
this.handle(req, res).catch((err) => {
|
|
1290
|
+
sendJson(res, 500, { error: { message: err instanceof Error ? err.message : String(err) } });
|
|
1291
|
+
});
|
|
1292
|
+
});
|
|
1293
|
+
await new Promise((resolve4, reject) => {
|
|
1294
|
+
this.server?.once("error", reject);
|
|
1295
|
+
this.server?.listen(this.options.publicPort, this.options.publicHost, () => {
|
|
1296
|
+
this.server?.off("error", reject);
|
|
1297
|
+
resolve4();
|
|
1298
|
+
});
|
|
1299
|
+
});
|
|
1300
|
+
}
|
|
1301
|
+
async stop() {
|
|
1302
|
+
this.clearIdleTimer();
|
|
1303
|
+
await this.stopChild();
|
|
1304
|
+
if (!this.server) return;
|
|
1305
|
+
const server = this.server;
|
|
1306
|
+
this.server = null;
|
|
1307
|
+
await new Promise((resolve4, reject) => {
|
|
1308
|
+
server.close((err) => err ? reject(err) : resolve4());
|
|
1309
|
+
});
|
|
1310
|
+
}
|
|
1311
|
+
async run() {
|
|
1312
|
+
await this.start();
|
|
1313
|
+
await new Promise((resolve4, reject) => {
|
|
1314
|
+
this.server?.once("error", reject);
|
|
1315
|
+
this.server?.once("close", resolve4);
|
|
1316
|
+
});
|
|
1317
|
+
}
|
|
1318
|
+
async handle(req, res) {
|
|
1319
|
+
const route = new URL(req.url || "/", "http://localhost").pathname;
|
|
1320
|
+
if (req.method === "OPTIONS") {
|
|
1321
|
+
sendJson(res, 204, {});
|
|
1322
|
+
return;
|
|
1323
|
+
}
|
|
1324
|
+
if (req.method === "GET" && route === "/health") {
|
|
1325
|
+
await this.handleHealth(res);
|
|
1326
|
+
return;
|
|
1327
|
+
}
|
|
1328
|
+
if (req.method === "POST" && COMPLETION_ROUTES.has(route)) {
|
|
1329
|
+
const receivedAt = Date.now();
|
|
1330
|
+
const body = await readRequestBody(req);
|
|
1331
|
+
this.queuedRequests += 1;
|
|
1332
|
+
const work = this.queue.then(() => this.processGeneration(route, body, res, receivedAt));
|
|
1333
|
+
this.queue = work.catch(() => void 0);
|
|
1334
|
+
try {
|
|
1335
|
+
await work;
|
|
1336
|
+
} finally {
|
|
1337
|
+
this.queuedRequests -= 1;
|
|
1338
|
+
this.scheduleIdleStop();
|
|
1339
|
+
}
|
|
1340
|
+
return;
|
|
1341
|
+
}
|
|
1342
|
+
sendJson(res, 404, { error: { message: "Not found" } });
|
|
1343
|
+
}
|
|
1344
|
+
async handleHealth(res) {
|
|
1345
|
+
if (!this.child) {
|
|
1346
|
+
sendJson(res, 200, {
|
|
1347
|
+
status: "ok",
|
|
1348
|
+
wrapper: "ok",
|
|
1349
|
+
model_status: "cold",
|
|
1350
|
+
restart_count: this.restartCount
|
|
1351
|
+
});
|
|
1352
|
+
return;
|
|
1353
|
+
}
|
|
1354
|
+
if (this.activeRequests > 0 || this.queuedRequests > 0) {
|
|
1355
|
+
sendJson(res, 200, {
|
|
1356
|
+
status: "ok",
|
|
1357
|
+
wrapper: "ok",
|
|
1358
|
+
model_status: "busy",
|
|
1359
|
+
restart_count: this.restartCount
|
|
1360
|
+
});
|
|
1361
|
+
return;
|
|
1362
|
+
}
|
|
1363
|
+
const health = await this.checkHealth();
|
|
1364
|
+
if (health.ok) {
|
|
1365
|
+
sendJson(res, 200, {
|
|
1366
|
+
status: "ok",
|
|
1367
|
+
wrapper: "ok",
|
|
1368
|
+
model_status: "warm",
|
|
1369
|
+
restart_count: this.restartCount,
|
|
1370
|
+
model: health.body
|
|
1371
|
+
});
|
|
1372
|
+
return;
|
|
1373
|
+
}
|
|
1374
|
+
await this.restartChild();
|
|
1375
|
+
sendJson(res, 200, {
|
|
1376
|
+
status: "ok",
|
|
1377
|
+
wrapper: "ok",
|
|
1378
|
+
model_status: "restarted",
|
|
1379
|
+
restart_count: this.restartCount
|
|
1380
|
+
});
|
|
1381
|
+
}
|
|
1382
|
+
async processGeneration(route, body, res, receivedAt) {
|
|
1383
|
+
const requestId = `ttreq-${randomUUID()}`;
|
|
1384
|
+
const startedAt = Date.now();
|
|
1385
|
+
const queuedMs = startedAt - receivedAt;
|
|
1386
|
+
this.activeRequests += 1;
|
|
1387
|
+
this.clearIdleTimer();
|
|
1388
|
+
let status = 500;
|
|
1389
|
+
let parsed = null;
|
|
1390
|
+
let failure = null;
|
|
1391
|
+
try {
|
|
1392
|
+
await this.ensureModelReady();
|
|
1393
|
+
const forwarded = await this.forwardGeneration(route, body);
|
|
1394
|
+
status = forwarded.status;
|
|
1395
|
+
parsed = forwarded.parsed;
|
|
1396
|
+
res.writeHead(status, {
|
|
1397
|
+
...copyResponseHeaders(forwarded.headers),
|
|
1398
|
+
"content-length": String(Buffer.byteLength(forwarded.body))
|
|
1399
|
+
});
|
|
1400
|
+
res.end(forwarded.body);
|
|
1401
|
+
this.requestsSinceStart += 1;
|
|
1402
|
+
failure = errorSummary(parsed, null);
|
|
1403
|
+
} catch (err) {
|
|
1404
|
+
failure = err instanceof Error ? err.message : String(err);
|
|
1405
|
+
status = 502;
|
|
1406
|
+
sendJson(res, status, { error: { message: failure } });
|
|
1407
|
+
} finally {
|
|
1408
|
+
const finishedAt = Date.now();
|
|
1409
|
+
this.activeRequests -= 1;
|
|
1410
|
+
this.writeLog({
|
|
1411
|
+
timestamp: new Date(finishedAt).toISOString(),
|
|
1412
|
+
request_id: requestId,
|
|
1413
|
+
route,
|
|
1414
|
+
model_target: {
|
|
1415
|
+
name: this.options.modelName,
|
|
1416
|
+
path: this.options.modelPath
|
|
1417
|
+
},
|
|
1418
|
+
request_bytes: body.byteLength,
|
|
1419
|
+
response_status: status,
|
|
1420
|
+
latency_ms: finishedAt - receivedAt,
|
|
1421
|
+
queued_ms: queuedMs,
|
|
1422
|
+
prompt_tokens: usageNumber(parsed, "prompt_tokens"),
|
|
1423
|
+
completion_tokens: usageNumber(parsed, "completion_tokens"),
|
|
1424
|
+
total_tokens: usageNumber(parsed, "total_tokens"),
|
|
1425
|
+
schema_validity: status < 400 ? true : status === 422 ? false : null,
|
|
1426
|
+
gate_field: this.options.gateField,
|
|
1427
|
+
gate_result: this.extractGateResult(parsed),
|
|
1428
|
+
restart_count: this.restartCount,
|
|
1429
|
+
error_summary: failure
|
|
1430
|
+
});
|
|
1431
|
+
}
|
|
1432
|
+
}
|
|
1433
|
+
async ensureModelReady() {
|
|
1434
|
+
if (this.child && this.restartAfterRequests > 0 && this.requestsSinceStart >= this.restartAfterRequests) {
|
|
1435
|
+
await this.restartChild();
|
|
1436
|
+
return;
|
|
1437
|
+
}
|
|
1438
|
+
if (this.child) {
|
|
1439
|
+
const health = await this.checkHealth();
|
|
1440
|
+
if (health.ok) return;
|
|
1441
|
+
await this.restartChild();
|
|
1442
|
+
return;
|
|
1443
|
+
}
|
|
1444
|
+
await this.startChild();
|
|
1445
|
+
}
|
|
1446
|
+
async startChild() {
|
|
1447
|
+
if (this.internalPort == null) {
|
|
1448
|
+
this.internalPort = await allocatePort("127.0.0.1");
|
|
1449
|
+
}
|
|
1450
|
+
const env = {
|
|
1451
|
+
...process.env,
|
|
1452
|
+
...this.options.env,
|
|
1453
|
+
TT_HOST: "127.0.0.1",
|
|
1454
|
+
TT_PORT: String(this.internalPort)
|
|
1455
|
+
};
|
|
1456
|
+
this.child = this.spawnModelServer(this.options.python, this.options.args, {
|
|
1457
|
+
stdio: "inherit",
|
|
1458
|
+
env
|
|
1459
|
+
});
|
|
1460
|
+
this.child.once("exit", () => {
|
|
1461
|
+
if (!this.stoppingChild) {
|
|
1462
|
+
this.child = null;
|
|
1463
|
+
this.requestsSinceStart = 0;
|
|
1464
|
+
}
|
|
1465
|
+
});
|
|
1466
|
+
try {
|
|
1467
|
+
await this.waitForHealth();
|
|
1468
|
+
this.scheduleIdleStop();
|
|
1469
|
+
} catch (err) {
|
|
1470
|
+
await this.stopChild();
|
|
1471
|
+
throw err;
|
|
1472
|
+
}
|
|
1473
|
+
}
|
|
1474
|
+
async restartChild() {
|
|
1475
|
+
await this.stopChild();
|
|
1476
|
+
this.restartCount += 1;
|
|
1477
|
+
this.requestsSinceStart = 0;
|
|
1478
|
+
await this.startChild();
|
|
1479
|
+
}
|
|
1480
|
+
async stopChild() {
|
|
1481
|
+
if (!this.child) return;
|
|
1482
|
+
const child = this.child;
|
|
1483
|
+
this.child = null;
|
|
1484
|
+
this.requestsSinceStart = 0;
|
|
1485
|
+
this.stoppingChild = true;
|
|
1486
|
+
try {
|
|
1487
|
+
child.kill("SIGTERM");
|
|
1488
|
+
await Promise.race([
|
|
1489
|
+
new Promise((resolve4) => child.once("exit", () => resolve4())),
|
|
1490
|
+
sleep(5e3).then(() => {
|
|
1491
|
+
if (child.exitCode == null && child.signalCode == null) child.kill("SIGKILL");
|
|
1492
|
+
})
|
|
1493
|
+
]);
|
|
1494
|
+
} finally {
|
|
1495
|
+
this.stoppingChild = false;
|
|
1496
|
+
}
|
|
1497
|
+
}
|
|
1498
|
+
async checkHealth() {
|
|
1499
|
+
if (this.internalPort == null) return { ok: false, body: null };
|
|
1500
|
+
const controller = new AbortController();
|
|
1501
|
+
const timeout = setTimeout(() => controller.abort(), 2e3);
|
|
1502
|
+
try {
|
|
1503
|
+
const response = await fetch(`http://127.0.0.1:${this.internalPort}/health`, {
|
|
1504
|
+
signal: controller.signal
|
|
1505
|
+
});
|
|
1506
|
+
const text = await response.text();
|
|
1507
|
+
return { ok: response.ok, body: tryJsonParse(text) };
|
|
1508
|
+
} catch {
|
|
1509
|
+
return { ok: false, body: null };
|
|
1510
|
+
} finally {
|
|
1511
|
+
clearTimeout(timeout);
|
|
1512
|
+
}
|
|
1513
|
+
}
|
|
1514
|
+
async waitForHealth() {
|
|
1515
|
+
const deadline = Date.now() + 12e4;
|
|
1516
|
+
while (Date.now() < deadline) {
|
|
1517
|
+
const health = await this.checkHealth();
|
|
1518
|
+
if (health.ok) return;
|
|
1519
|
+
if (!this.child) break;
|
|
1520
|
+
await sleep(50);
|
|
1521
|
+
}
|
|
1522
|
+
throw new Error("Model server did not become healthy within 120 seconds");
|
|
1523
|
+
}
|
|
1524
|
+
async forwardGeneration(route, body) {
|
|
1525
|
+
if (this.internalPort == null) throw new Error("Internal server port is not allocated");
|
|
1526
|
+
const response = await fetch(`http://127.0.0.1:${this.internalPort}${route}`, {
|
|
1527
|
+
method: "POST",
|
|
1528
|
+
headers: { "content-type": "application/json" },
|
|
1529
|
+
body
|
|
1530
|
+
});
|
|
1531
|
+
const text = await response.text();
|
|
1532
|
+
return {
|
|
1533
|
+
status: response.status,
|
|
1534
|
+
headers: response.headers,
|
|
1535
|
+
body: text,
|
|
1536
|
+
parsed: tryJsonParse(text)
|
|
1537
|
+
};
|
|
1538
|
+
}
|
|
1539
|
+
extractGateResult(response) {
|
|
1540
|
+
const content = extractAssistantContent(response);
|
|
1541
|
+
if (!content) return null;
|
|
1542
|
+
const parsedContent = tryJsonParse(content);
|
|
1543
|
+
return readPath(parsedContent, this.options.gateField);
|
|
1544
|
+
}
|
|
1545
|
+
scheduleIdleStop() {
|
|
1546
|
+
this.clearIdleTimer();
|
|
1547
|
+
if (!this.child || this.activeRequests > 0 || this.queuedRequests > 0) return;
|
|
1548
|
+
this.idleTimer = setTimeout(() => {
|
|
1549
|
+
this.stopChild().catch((err) => {
|
|
1550
|
+
this.writeDiagnosticError(err);
|
|
1551
|
+
});
|
|
1552
|
+
}, this.idleTimeoutMs);
|
|
1553
|
+
}
|
|
1554
|
+
clearIdleTimer() {
|
|
1555
|
+
if (!this.idleTimer) return;
|
|
1556
|
+
clearTimeout(this.idleTimer);
|
|
1557
|
+
this.idleTimer = null;
|
|
1558
|
+
}
|
|
1559
|
+
writeLog(record) {
|
|
1560
|
+
if (this.logRecordOverride) {
|
|
1561
|
+
this.logRecordOverride(record);
|
|
1562
|
+
return;
|
|
1563
|
+
}
|
|
1564
|
+
const line = `${JSON.stringify(record)}
|
|
1565
|
+
`;
|
|
1566
|
+
if (this.options.logFile) {
|
|
1567
|
+
mkdirSync2(dirname(this.options.logFile), { recursive: true });
|
|
1568
|
+
appendFileSync(this.options.logFile, line, "utf8");
|
|
1569
|
+
return;
|
|
1570
|
+
}
|
|
1571
|
+
process.stderr.write(line);
|
|
1572
|
+
}
|
|
1573
|
+
writeDiagnosticError(err) {
|
|
1574
|
+
const message = err instanceof Error ? err.message : String(err);
|
|
1575
|
+
process.stderr.write(`${JSON.stringify({ timestamp: (/* @__PURE__ */ new Date()).toISOString(), error: message })}
|
|
1576
|
+
`);
|
|
1577
|
+
}
|
|
1578
|
+
};
|
|
1579
|
+
async function startManagedModelServer(options) {
|
|
1580
|
+
const server = new ManagedModelServer(options);
|
|
1581
|
+
const stop = () => {
|
|
1582
|
+
server.stop().catch(() => void 0);
|
|
1583
|
+
};
|
|
1584
|
+
process.once("SIGINT", stop);
|
|
1585
|
+
process.once("SIGTERM", stop);
|
|
1586
|
+
try {
|
|
1587
|
+
await server.run();
|
|
1588
|
+
} finally {
|
|
1589
|
+
process.off("SIGINT", stop);
|
|
1590
|
+
process.off("SIGTERM", stop);
|
|
1591
|
+
await server.stop();
|
|
1592
|
+
}
|
|
1593
|
+
}
|
|
1594
|
+
|
|
1077
1595
|
// src/commands/models.ts
|
|
1078
1596
|
var REFERENCE_SERVER_SCRIPT = String.raw`#!/usr/bin/env python3
|
|
1079
1597
|
import json
|
|
@@ -1585,7 +2103,7 @@ async function downloadUrlToFile(url, outputPath) {
|
|
|
1585
2103
|
if (!res.body) {
|
|
1586
2104
|
throw new Error("Download failed: response body was empty");
|
|
1587
2105
|
}
|
|
1588
|
-
|
|
2106
|
+
mkdirSync3(dirname2(outputPath), { recursive: true });
|
|
1589
2107
|
const file = createWriteStream(outputPath);
|
|
1590
2108
|
const contentLength = parseContentLength(res.headers.get("content-length"));
|
|
1591
2109
|
const progress = createDownloadProgress(
|
|
@@ -1713,7 +2231,7 @@ function directoryHasFiles(path) {
|
|
|
1713
2231
|
function extractArchive(archivePath, targetDir, force) {
|
|
1714
2232
|
if (directoryHasFiles(targetDir) && !force) return targetDir;
|
|
1715
2233
|
if (existsSync4(targetDir)) rmSync(targetDir, { recursive: true, force: true });
|
|
1716
|
-
|
|
2234
|
+
mkdirSync3(targetDir, { recursive: true });
|
|
1717
2235
|
execFileSync("tar", ["-xzf", archivePath, "-C", targetDir], { stdio: "inherit" });
|
|
1718
2236
|
return targetDir;
|
|
1719
2237
|
}
|
|
@@ -1761,7 +2279,7 @@ function loadJsonSchemaForServe(schemaPath) {
|
|
|
1761
2279
|
}
|
|
1762
2280
|
function writeReferenceServerScript(cacheDir) {
|
|
1763
2281
|
const scriptDir = join2(cacheDir, "_server");
|
|
1764
|
-
|
|
2282
|
+
mkdirSync3(scriptDir, { recursive: true });
|
|
1765
2283
|
const scriptPath = join2(scriptDir, "openai_reference_server.py");
|
|
1766
2284
|
writeFileSync3(scriptPath, REFERENCE_SERVER_SCRIPT, "utf8");
|
|
1767
2285
|
return scriptPath;
|
|
@@ -1808,16 +2326,34 @@ function quoteCommandPart(part) {
|
|
|
1808
2326
|
if (/^[a-zA-Z0-9_./:=@-]+$/.test(part)) return part;
|
|
1809
2327
|
return `'${part.replace(/'/g, "'\\''")}'`;
|
|
1810
2328
|
}
|
|
1811
|
-
function
|
|
2329
|
+
function parseServeInteger(value, label) {
|
|
2330
|
+
const parsed = Number(value);
|
|
2331
|
+
if (!Number.isInteger(parsed) || parsed < 0) {
|
|
2332
|
+
throw new Error(`${label} must be a non-negative integer.`);
|
|
2333
|
+
}
|
|
2334
|
+
return parsed;
|
|
2335
|
+
}
|
|
2336
|
+
function printServeCommand(command, args, env, managed) {
|
|
1812
2337
|
const envKeys = Object.keys(env).sort();
|
|
1813
2338
|
const commandLine = [command, ...args].map(quoteCommandPart).join(" ");
|
|
1814
2339
|
if (isJsonMode()) {
|
|
1815
|
-
return printJson({ command, args, env_keys: envKeys, command_line: commandLine });
|
|
2340
|
+
return printJson({ command, args, env_keys: envKeys, command_line: commandLine, managed });
|
|
1816
2341
|
}
|
|
1817
|
-
|
|
2342
|
+
const fields = [
|
|
1818
2343
|
["Command", commandLine],
|
|
1819
2344
|
["Environment", envKeys.join(", ")]
|
|
1820
|
-
]
|
|
2345
|
+
];
|
|
2346
|
+
if (managed?.enabled) {
|
|
2347
|
+
fields.push(
|
|
2348
|
+
["Managed", "enabled"],
|
|
2349
|
+
["Wrapper", `http://${managed.host}:${managed.port}`],
|
|
2350
|
+
["Idle timeout", `${managed.idle_timeout_seconds}s`],
|
|
2351
|
+
["Restart after", managed.restart_after_requests === 0 ? "disabled" : `${managed.restart_after_requests} requests`],
|
|
2352
|
+
["Gate field", managed.gate_field],
|
|
2353
|
+
["Log file", managed.log_file]
|
|
2354
|
+
);
|
|
2355
|
+
}
|
|
2356
|
+
printDetail(fields);
|
|
1821
2357
|
}
|
|
1822
2358
|
function registerModelsCommands(parent) {
|
|
1823
2359
|
const models = parent.command("models").description("Manage fine-tuned models");
|
|
@@ -1921,7 +2457,7 @@ function registerModelsCommands(parent) {
|
|
|
1921
2457
|
rmSync(venvDir, { recursive: true, force: true });
|
|
1922
2458
|
}
|
|
1923
2459
|
if (!existsSync4(venvPython)) {
|
|
1924
|
-
|
|
2460
|
+
mkdirSync3(cacheDir, { recursive: true });
|
|
1925
2461
|
execFileSync(commands[0][0], commands[0].slice(1), { stdio: "inherit" });
|
|
1926
2462
|
}
|
|
1927
2463
|
execFileSync(commands[1][0], commands[1].slice(1), { stdio: "inherit" });
|
|
@@ -1938,10 +2474,10 @@ function registerModelsCommands(parent) {
|
|
|
1938
2474
|
printSuccess(`Serving runtime ready at ${venvDir}`);
|
|
1939
2475
|
console.log(`Use it with: tt models serve <model-id> --spec ${DEFAULT_SPEC_FILE}`);
|
|
1940
2476
|
});
|
|
1941
|
-
models.command("serve").description("Serve a downloaded model with an OpenAI-compatible local API").argument("<target>", "Model ID/prefix, model directory, or .tar.gz artifact").option("--spec <path>", "Behavior spec JSON to apply as the default system prompt").option("--no-spec-prompt", "Do not auto-apply a behavior spec system prompt").option("--host <host>", "Host to bind", "127.0.0.1").option("--port <port>", "Port to bind", "8000").option("--python <path>", "Python executable").option("--cache-dir <path>", "Cache directory for downloaded/extracted models").option("--device <device>", "Inference device: auto, cpu, cuda, or mps", "auto").option("--force-download", "Re-download and re-extract model artifacts").option("--max-tokens <n>", "Default max completion tokens", "512").option("--temperature <n>", "Default sampling temperature", "0.7").option("--json-schema <path>", "JSON Schema file to enforce by default for chat completions").option("--json-repair-attempts <n>", "Number of JSON/schema repair retries", "1").option("--trust-remote-code", "Pass trust_remote_code=True to transformers").option("--print-command", "Print the underlying Python command without starting it").action(async (target, cmdOpts) => {
|
|
2477
|
+
models.command("serve").description("Serve a downloaded model with an OpenAI-compatible local API").argument("<target>", "Model ID/prefix, model directory, or .tar.gz artifact").option("--spec <path>", "Behavior spec JSON to apply as the default system prompt").option("--no-spec-prompt", "Do not auto-apply a behavior spec system prompt").option("--host <host>", "Host to bind", "127.0.0.1").option("--port <port>", "Port to bind", "8000").option("--python <path>", "Python executable").option("--cache-dir <path>", "Cache directory for downloaded/extracted models").option("--device <device>", "Inference device: auto, cpu, cuda, or mps", "auto").option("--force-download", "Re-download and re-extract model artifacts").option("--max-tokens <n>", "Default max completion tokens", "512").option("--temperature <n>", "Default sampling temperature", "0.7").option("--json-schema <path>", "JSON Schema file to enforce by default for chat completions").option("--json-repair-attempts <n>", "Number of JSON/schema repair retries", "1").option("--trust-remote-code", "Pass trust_remote_code=True to transformers").option("--managed", "Run a local lifecycle manager in front of the model server").option("--idle-timeout <seconds>", "Managed mode idle timeout before stopping the model", "300").option("--restart-after-requests <n>", "Managed mode restart threshold; 0 disables request-count restarts", "100").option("--gate-field <field>", "Response JSON field to log as the managed serving gate result", "should_process").option("--log-file <path>", "Write managed serving JSONL request logs to a file").option("--print-command", "Print the underlying Python command without starting it").action(async (target, cmdOpts) => {
|
|
1942
2478
|
const opts = parent.opts();
|
|
1943
2479
|
const cacheDir = resolve2(cmdOpts.cacheDir || defaultCacheDir());
|
|
1944
|
-
|
|
2480
|
+
mkdirSync3(cacheDir, { recursive: true });
|
|
1945
2481
|
const python = resolveServePython(cacheDir, cmdOpts.python);
|
|
1946
2482
|
if (!cmdOpts.printCommand) ensureServingRuntime(python, cacheDir);
|
|
1947
2483
|
const serveTarget = await resolveServeTarget(
|
|
@@ -1976,9 +2512,26 @@ function registerModelsCommands(parent) {
|
|
|
1976
2512
|
if (systemPrompt) env.TT_SYSTEM_PROMPT = systemPrompt;
|
|
1977
2513
|
if (cmdOpts.jsonSchema) env.TT_JSON_SCHEMA = loadJsonSchemaForServe(cmdOpts.jsonSchema);
|
|
1978
2514
|
const args = [serverScript];
|
|
1979
|
-
|
|
2515
|
+
const idleTimeoutSeconds = parseServeInteger(cmdOpts.idleTimeout, "--idle-timeout");
|
|
2516
|
+
const restartAfterRequests = parseServeInteger(
|
|
2517
|
+
cmdOpts.restartAfterRequests,
|
|
2518
|
+
"--restart-after-requests"
|
|
2519
|
+
);
|
|
2520
|
+
const publicPort = parseServeInteger(cmdOpts.port, "--port");
|
|
2521
|
+
const managedConfig = cmdOpts.managed ? {
|
|
2522
|
+
enabled: true,
|
|
2523
|
+
host: String(cmdOpts.host),
|
|
2524
|
+
port: publicPort,
|
|
2525
|
+
idle_timeout_seconds: idleTimeoutSeconds,
|
|
2526
|
+
restart_after_requests: restartAfterRequests,
|
|
2527
|
+
gate_field: String(cmdOpts.gateField),
|
|
2528
|
+
log_file: cmdOpts.logFile ? resolve2(cmdOpts.logFile) : void 0
|
|
2529
|
+
} : void 0;
|
|
2530
|
+
if (cmdOpts.printCommand) return printServeCommand(python, args, env, managedConfig);
|
|
1980
2531
|
if (!isJsonMode()) {
|
|
1981
|
-
printSuccess(
|
|
2532
|
+
printSuccess(
|
|
2533
|
+
cmdOpts.managed ? `Serving ${serveTarget.modelName} through a managed local wrapper` : `Serving ${serveTarget.modelName} from ${serveTarget.modelPath}`
|
|
2534
|
+
);
|
|
1982
2535
|
if (systemPrompt && specPath) {
|
|
1983
2536
|
printSuccess(`Behavior spec prompt loaded from ${specPath}`);
|
|
1984
2537
|
} else if (cmdOpts.specPrompt !== false) {
|
|
@@ -1989,9 +2542,32 @@ function registerModelsCommands(parent) {
|
|
|
1989
2542
|
console.log(`OpenAI-compatible endpoint: http://${cmdOpts.host}:${cmdOpts.port}/v1/chat/completions`);
|
|
1990
2543
|
console.log(`Health check: http://${cmdOpts.host}:${cmdOpts.port}/health`);
|
|
1991
2544
|
console.log(`Python runtime: ${python}`);
|
|
2545
|
+
if (cmdOpts.managed) {
|
|
2546
|
+
console.log(`Managed idle timeout: ${idleTimeoutSeconds}s`);
|
|
2547
|
+
console.log(
|
|
2548
|
+
`Managed restart threshold: ${restartAfterRequests === 0 ? "disabled" : `${restartAfterRequests} requests`}`
|
|
2549
|
+
);
|
|
2550
|
+
console.log(`Managed gate field: ${cmdOpts.gateField}`);
|
|
2551
|
+
}
|
|
2552
|
+
}
|
|
2553
|
+
if (cmdOpts.managed) {
|
|
2554
|
+
await startManagedModelServer({
|
|
2555
|
+
publicHost: String(cmdOpts.host),
|
|
2556
|
+
publicPort,
|
|
2557
|
+
python,
|
|
2558
|
+
args,
|
|
2559
|
+
env,
|
|
2560
|
+
modelName: serveTarget.modelName,
|
|
2561
|
+
modelPath: serveTarget.modelPath,
|
|
2562
|
+
idleTimeoutSeconds,
|
|
2563
|
+
restartAfterRequests,
|
|
2564
|
+
gateField: String(cmdOpts.gateField),
|
|
2565
|
+
logFile: cmdOpts.logFile ? resolve2(cmdOpts.logFile) : void 0
|
|
2566
|
+
});
|
|
2567
|
+
return;
|
|
1992
2568
|
}
|
|
1993
2569
|
await new Promise((resolvePromise, reject) => {
|
|
1994
|
-
const child =
|
|
2570
|
+
const child = spawn2(python, args, {
|
|
1995
2571
|
stdio: "inherit",
|
|
1996
2572
|
env: { ...process.env, ...env }
|
|
1997
2573
|
});
|
|
@@ -2305,7 +2881,7 @@ function registerPushCommand(parent) {
|
|
|
2305
2881
|
|
|
2306
2882
|
// src/index.ts
|
|
2307
2883
|
var program = new Command();
|
|
2308
|
-
program.name("tt").description("Tuned Tensor CLI \u2014 fine-tune and evaluate LLMs").version("0.4.
|
|
2884
|
+
program.name("tt").description("Tuned Tensor CLI \u2014 fine-tune and evaluate LLMs").version("0.4.15").option("-k, --api-key <key>", "API key (overrides stored key)").option(
|
|
2309
2885
|
"-u, --base-url <url>",
|
|
2310
2886
|
"API base URL (default: https://tunedtensor.com)"
|
|
2311
2887
|
).option("--json", "Output raw JSON").option("--no-color", "Disable colors").hook("preAction", (_thisCommand, actionCommand) => {
|