@infernetprotocol/cli 0.1.52 → 0.1.53
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/commands/start.js +36 -5
- package/lib/vllm.js +14 -3
- package/package.json +7 -7
package/commands/start.js
CHANGED
|
@@ -48,7 +48,7 @@ import { executeChatJob, failChatJob, shutdownEngine } from '../lib/chat-executo
|
|
|
48
48
|
import { gatherCoarseSpecs } from './register.js';
|
|
49
49
|
import { checkModelFits, formatFitFailure } from '../lib/model-fit.js';
|
|
50
50
|
import { downloadHfModel, resolveHfToken } from '../lib/hf-model.js';
|
|
51
|
-
import { startVllmServe, vllmInstalled, waitForVllmModel, extractVllmError } from '../lib/vllm.js';
|
|
51
|
+
import { startVllmServe, stopVllmServe, vllmInstalled, waitForVllmModel, extractVllmError } from '../lib/vllm.js';
|
|
52
52
|
import { detectGpus, detectHost } from '@infernetprotocol/gpu';
|
|
53
53
|
import { getOrCreateModelKey, getModelPublicKeys } from '../lib/model-key.js';
|
|
54
54
|
import { pullLatestBinary } from './upgrade.js';
|
|
@@ -643,24 +643,55 @@ async function runDaemon(args, ctx) {
|
|
|
643
643
|
// model name so chat requests + served_models match, but the
|
|
644
644
|
// engine underneath is vLLM.
|
|
645
645
|
const token = process.env.HF_TOKEN || (await resolveHfToken().catch(() => null));
|
|
646
|
+
// Step 1/3 — download weights.
|
|
647
|
+
await client.updateCommandProgress(id, {
|
|
648
|
+
phase: 'download', step: 1, steps: 3, status: 'downloading weights'
|
|
649
|
+
}).catch(() => {});
|
|
646
650
|
const localPath = await downloadHfModel(hfRepo, token);
|
|
647
651
|
// Don't report "completed" on spawn — wait until vLLM has
|
|
648
652
|
// mapped the weights and /v1/models actually lists the model.
|
|
649
653
|
// If it never comes up (OOM, unsupported arch, gated repo),
|
|
650
654
|
// fail the command WITH the log root cause instead of lying.
|
|
655
|
+
// Step 2/3 — start the engine + wait for it to serve.
|
|
656
|
+
await client.updateCommandProgress(id, {
|
|
657
|
+
phase: 'serve', step: 2, steps: 3, status: 'starting vLLM engine', pct: 0
|
|
658
|
+
}).catch(() => {});
|
|
651
659
|
const serve = await startVllmServe({ source: hfRepo, servedName: modelName, token });
|
|
652
660
|
process.stdout.write(`[${new Date().toISOString()}] ${modelName} (${hfRepo}) downloaded; waiting for vLLM to load weights…\n`);
|
|
653
|
-
//
|
|
654
|
-
//
|
|
655
|
-
|
|
656
|
-
|
|
661
|
+
// Auto-cancel ceiling: give up after VLLM_INSTALL_TIMEOUT_S
|
|
662
|
+
// (default 10 min). Each tick reports progress AND checks
|
|
663
|
+
// whether the owner cancelled (the progress response echoes
|
|
664
|
+
// the command status) so a stuck install can be stopped.
|
|
665
|
+
const vllmWaitMs = Number(process.env.VLLM_INSTALL_TIMEOUT_S || 600) * 1000;
|
|
666
|
+
const up = await waitForVllmModel(modelName, {
|
|
667
|
+
pid: serve.pid,
|
|
668
|
+
timeoutMs: vllmWaitMs,
|
|
669
|
+
onTick: async ({ elapsedMs, timeoutMs }) => {
|
|
670
|
+
const pct = Math.min(99, Math.round((elapsedMs / timeoutMs) * 100));
|
|
671
|
+
const resp = await client.updateCommandProgress(id, {
|
|
672
|
+
phase: 'serve', step: 2, steps: 3, status: 'starting vLLM engine', pct
|
|
673
|
+
}).catch(() => null);
|
|
674
|
+
if (resp?.status === 'cancelled') return 'cancel';
|
|
675
|
+
}
|
|
676
|
+
});
|
|
677
|
+
if (up.cancelled) {
|
|
678
|
+
await stopVllmServe().catch(() => {});
|
|
679
|
+
invalidateSpecsCache();
|
|
680
|
+
process.stdout.write(`[${new Date().toISOString()}] ${modelName} install cancelled — stopped vLLM\n`);
|
|
681
|
+
return; // status is already 'cancelled'; don't complete
|
|
682
|
+
}
|
|
657
683
|
if (!up.serving) {
|
|
684
|
+
await stopVllmServe().catch(() => {});
|
|
658
685
|
const tail = await extractVllmError();
|
|
659
686
|
throw new Error(
|
|
660
687
|
`vLLM failed to serve ${modelName} (${hfRepo}): ${up.reason}.` +
|
|
661
688
|
(tail ? `\n--- vllm.log (root cause) ---\n${tail}` : ' (no vllm.log output)')
|
|
662
689
|
);
|
|
663
690
|
}
|
|
691
|
+
// Step 3/3 — serving.
|
|
692
|
+
await client.updateCommandProgress(id, {
|
|
693
|
+
phase: 'serve', step: 3, steps: 3, status: 'serving', pct: 100
|
|
694
|
+
}).catch(() => {});
|
|
664
695
|
invalidateSpecsCache();
|
|
665
696
|
result = {
|
|
666
697
|
model: modelName,
|
package/lib/vllm.js
CHANGED
|
@@ -150,8 +150,9 @@ export async function extractVllmError({ headLines = 16, tailLines = 30, fallbac
|
|
|
150
150
|
* appears, bails early if the process dies, and times out otherwise.
|
|
151
151
|
* A 9B model maps to GPU in ~1-2 min, so the default budget is generous.
|
|
152
152
|
*/
|
|
153
|
-
export async function waitForVllmModel(servedName, { pid = null, timeoutMs = 300_000, intervalMs = 3000 } = {}) {
|
|
154
|
-
const
|
|
153
|
+
export async function waitForVllmModel(servedName, { pid = null, timeoutMs = 300_000, intervalMs = 3000, onTick = null } = {}) {
|
|
154
|
+
const start = Date.now();
|
|
155
|
+
const deadline = start + timeoutMs;
|
|
155
156
|
while (Date.now() < deadline) {
|
|
156
157
|
if (pid && !isPidAlive(pid)) {
|
|
157
158
|
return { serving: false, reason: 'vLLM process exited during startup (check the log tail below)' };
|
|
@@ -160,9 +161,19 @@ export async function waitForVllmModel(servedName, { pid = null, timeoutMs = 300
|
|
|
160
161
|
if (servedName ? models.includes(servedName) : models.length > 0) {
|
|
161
162
|
return { serving: true, models };
|
|
162
163
|
}
|
|
164
|
+
// Per-tick hook: report progress + let the caller abort (user cancel).
|
|
165
|
+
// Returning 'cancel' (or a truthy .cancel) stops the wait immediately.
|
|
166
|
+
if (onTick) {
|
|
167
|
+
try {
|
|
168
|
+
const r = await onTick({ elapsedMs: Date.now() - start, timeoutMs });
|
|
169
|
+
if (r === 'cancel' || r?.cancel) {
|
|
170
|
+
return { serving: false, cancelled: true, reason: 'cancelled by user' };
|
|
171
|
+
}
|
|
172
|
+
} catch { /* onTick is best-effort */ }
|
|
173
|
+
}
|
|
163
174
|
await sleep(intervalMs);
|
|
164
175
|
}
|
|
165
|
-
return { serving: false, reason: `timed out after ${Math.round(timeoutMs / 1000)}s waiting for /v1/models to list ${servedName}` };
|
|
176
|
+
return { serving: false, timedOut: true, reason: `timed out after ${Math.round(timeoutMs / 1000)}s waiting for /v1/models to list ${servedName}` };
|
|
166
177
|
}
|
|
167
178
|
|
|
168
179
|
/** True when the tracked serve pid is alive AND /v1/models answers. */
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@infernetprotocol/cli",
|
|
3
|
-
"version": "0.1.
|
|
3
|
+
"version": "0.1.53",
|
|
4
4
|
"description": "GPU-node CLI / daemon for the Infernet Protocol — register a GPU server with a Supabase control plane and start earning.",
|
|
5
5
|
"license": "MIT",
|
|
6
6
|
"type": "module",
|
|
@@ -30,12 +30,12 @@
|
|
|
30
30
|
"ora": "^9.4.0",
|
|
31
31
|
"react": "^17.0.2",
|
|
32
32
|
"react-blessed": "^0.7.2",
|
|
33
|
-
"@infernetprotocol/
|
|
34
|
-
"@infernetprotocol/
|
|
35
|
-
"@infernetprotocol/
|
|
36
|
-
"@infernetprotocol/
|
|
37
|
-
"@infernetprotocol/
|
|
38
|
-
"@infernetprotocol/
|
|
33
|
+
"@infernetprotocol/auth": "0.1.53",
|
|
34
|
+
"@infernetprotocol/deploy-providers": "0.1.53",
|
|
35
|
+
"@infernetprotocol/engine": "0.1.53",
|
|
36
|
+
"@infernetprotocol/gpu": "0.1.53",
|
|
37
|
+
"@infernetprotocol/config": "0.1.53",
|
|
38
|
+
"@infernetprotocol/rpc-adapter": "0.1.53"
|
|
39
39
|
},
|
|
40
40
|
"repository": {
|
|
41
41
|
"type": "git",
|