@infernetprotocol/cli 0.1.37 → 0.1.39

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.
@@ -32,6 +32,7 @@ Usage:
32
32
  infernet inference serve --model <hf-id> [flags]
33
33
  infernet inference status
34
34
  infernet inference stop
35
+ infernet inference list
35
36
 
36
37
  Flags (serve):
37
38
  --model <hf-id> Required. e.g. meta-llama/Llama-3.1-70B-Instruct
@@ -171,6 +172,30 @@ async function cmdStatus() {
171
172
  return 0;
172
173
  }
173
174
 
175
+ async function cmdList() {
176
+ const cfg = await loadConfig().catch(() => ({}));
177
+ const controlPlaneUrl = cfg?.controlPlane?.url ?? process.env.NEXT_PUBLIC_APP_URL ?? "https://infernetprotocol.com";
178
+ const res = await fetch(`${controlPlaneUrl}/api/v1/petals/swarm`);
179
+ if (!res.ok) {
180
+ process.stderr.write(`error: HTTP ${res.status} from ${controlPlaneUrl}\n`);
181
+ return 1;
182
+ }
183
+ const { data } = await res.json();
184
+ if (!data?.models?.length) {
185
+ process.stdout.write(`(no Petals swarms active right now)\n`);
186
+ process.stdout.write(`Be the first — \`infernet inference serve --backend petals --model <hf-id>\`\n`);
187
+ return 0;
188
+ }
189
+ process.stdout.write(`\nPetals swarms — ${data.total_models} models · ${data.total_nodes} nodes serving\n\n`);
190
+ process.stdout.write(` ${"NODES".padEnd(7)} MODEL\n`);
191
+ process.stdout.write(` ${"─────".padEnd(7)} ─────\n`);
192
+ for (const m of data.models) {
193
+ process.stdout.write(` ${String(m.node_count).padEnd(7)} ${m.model}\n`);
194
+ }
195
+ process.stdout.write(`\nQuery as a user: tick \"Distribute across all nodes\" on /chat and pick a model above.\n\n`);
196
+ return 0;
197
+ }
198
+
174
199
  async function cmdStop() {
175
200
  const state = await readState();
176
201
  if (!state?.pid) {
@@ -197,6 +222,7 @@ export default async function inference(args) {
197
222
  case "serve": case "start": return cmdServe(args);
198
223
  case "status": return cmdStatus();
199
224
  case "stop": return cmdStop();
225
+ case "list": case "ls": return cmdList();
200
226
  default:
201
227
  process.stderr.write(sub ? `unknown subcommand: ${sub}\n\n` : "error: missing subcommand\n\n");
202
228
  process.stderr.write(HELP);
@@ -204,10 +204,33 @@ export async function gatherCoarseSpecs() {
204
204
  model: typeof g.model === 'string' ? g.model.slice(0, 64) : null
205
205
  })),
206
206
  interconnects: summarizeInterconnects(interconnects),
207
- served_models
207
+ served_models,
208
+ // IPIP-0031: models this node is contributing to a Petals swarm.
209
+ // Empty array if the operator hasn't run `infernet inference serve`.
210
+ // Read from local state file written by inference.js.
211
+ petals_models: await readPetalsModels()
208
212
  };
209
213
  }
210
214
 
215
+ async function readPetalsModels() {
216
+ try {
217
+ const fsp = await import('node:fs/promises');
218
+ const path = await import('node:path');
219
+ const statePath = path.join(process.env.HOME ?? '/tmp', '.infernet', 'inference', 'state.json');
220
+ const raw = await fsp.readFile(statePath, 'utf8');
221
+ const state = JSON.parse(raw);
222
+ // Only advertise if the server is still alive.
223
+ if (state.pid) {
224
+ try { process.kill(state.pid, 0); } // 0 = check liveness
225
+ catch { return []; } // pid stale
226
+ } else { return []; }
227
+ if (state.backend === 'petals' && typeof state.model === 'string') {
228
+ return [state.model];
229
+ }
230
+ return [];
231
+ } catch { return []; }
232
+ }
233
+
211
234
  export default async function register(args, ctx) {
212
235
  if (args.has('help') || args.has('h')) {
213
236
  process.stdout.write(HELP);
package/commands/start.js CHANGED
@@ -776,6 +776,78 @@ async function runDaemon(args, ctx) {
776
776
  };
777
777
  }
778
778
 
779
+ /**
780
+ * IPIP-0031: distributed inference proxy. Accepts a chat request,
781
+ * spawns a Python Petals client subprocess, streams its JSON-line
782
+ * output back as SSE.
783
+ *
784
+ * POST /v1/petals/inference
785
+ * body: { model, messages, max_tokens?, temperature? }
786
+ * response: text/event-stream with token / done / error frames
787
+ */
788
+ async function handlePetalsInference(req, res) {
789
+ if (req.method !== 'POST' || req.url !== '/v1/petals/inference') return false;
790
+
791
+ const fsp = await import('node:fs/promises');
792
+ const path = await import('node:path');
793
+ const os = await import('node:os');
794
+ const { PETALS_CLIENT_PY } = await import('../lib/inference/petals-client.py.js');
795
+
796
+ // Read JSON body
797
+ let body = '';
798
+ await new Promise((resolve, reject) => {
799
+ req.on('data', (chunk) => { body += chunk.toString(); });
800
+ req.on('end', resolve);
801
+ req.on('error', reject);
802
+ });
803
+ let payload;
804
+ try { payload = JSON.parse(body); } catch {
805
+ res.writeHead(400, { 'content-type': 'text/plain' });
806
+ res.end('bad json\n');
807
+ return true;
808
+ }
809
+
810
+ // Drop the bundled Python client to a temp file and spawn it.
811
+ const tmp = path.join(os.tmpdir(), `petals-${process.pid}-${Date.now()}.py`);
812
+ await fsp.writeFile(tmp, PETALS_CLIENT_PY);
813
+
814
+ res.writeHead(200, {
815
+ 'content-type': 'text/event-stream',
816
+ 'cache-control': 'no-cache',
817
+ 'x-accel-buffering': 'no'
818
+ });
819
+
820
+ const child = spawn('python3', [tmp], { stdio: ['pipe', 'pipe', 'pipe'] });
821
+ child.stdin.write(JSON.stringify(payload));
822
+ child.stdin.end();
823
+
824
+ let stdoutBuf = '';
825
+ child.stdout.on('data', (chunk) => {
826
+ stdoutBuf += chunk.toString();
827
+ let nl;
828
+ while ((nl = stdoutBuf.indexOf('\n')) >= 0) {
829
+ const line = stdoutBuf.slice(0, nl).trim();
830
+ stdoutBuf = stdoutBuf.slice(nl + 1);
831
+ if (!line) continue;
832
+ try {
833
+ const ev = JSON.parse(line);
834
+ res.write(`event: ${ev.event ?? 'data'}\ndata: ${JSON.stringify(ev)}\n\n`);
835
+ } catch {
836
+ res.write(`event: log\ndata: ${JSON.stringify({ raw: line })}\n\n`);
837
+ }
838
+ }
839
+ });
840
+ child.stderr.on('data', (chunk) => {
841
+ res.write(`event: log\ndata: ${JSON.stringify({ stderr: chunk.toString() })}\n\n`);
842
+ });
843
+ child.on('exit', async (code) => {
844
+ res.write(`event: exit\ndata: ${JSON.stringify({ code })}\n\n`);
845
+ res.end();
846
+ try { await fsp.unlink(tmp); } catch { /* ignore */ }
847
+ });
848
+ return true;
849
+ }
850
+
779
851
  /**
780
852
  * IPIP-0030: shard-hosting routes layered on top of /healthz.
781
853
  * Returns true if the request matched and was handled.
@@ -1063,6 +1135,15 @@ async function runDaemon(args, ctx) {
1063
1135
  return;
1064
1136
  }
1065
1137
 
1138
+ try {
1139
+ const petalsHandled = await handlePetalsInference(req, res);
1140
+ if (petalsHandled) return;
1141
+ } catch (err) {
1142
+ res.writeHead(500, { 'content-type': 'text/plain' });
1143
+ res.end(`petals http error: ${err?.message ?? err}\n`);
1144
+ return;
1145
+ }
1146
+
1066
1147
  res.writeHead(404, { 'content-type': 'text/plain' });
1067
1148
  res.end('not found\n');
1068
1149
  });
@@ -0,0 +1,98 @@
1
+ /**
2
+ * Bundled Python petals client. Spawned by the daemon when it accepts
3
+ * an /v1/petals/inference POST. Fans out across the public Petals
4
+ * swarm using AutoDistributedModelForCausalLM and streams JSON lines
5
+ * (one per token) to stdout. The daemon converts those to SSE frames.
6
+ */
7
+
8
+ export const PETALS_CLIENT_PY = `#!/usr/bin/env python3
9
+ """Petals client — fans out across the swarm + streams tokens.
10
+
11
+ Reads: JSON request from stdin: { model, messages, max_tokens, temperature }
12
+ Writes: JSON lines to stdout: { event: 'token' | 'done' | 'error', ... }
13
+
14
+ Each output line is a single JSON object terminated with \\n so the
15
+ parent process can parse them line-by-line.
16
+ """
17
+ import json
18
+ import sys
19
+ import time
20
+
21
+ def emit(obj):
22
+ sys.stdout.write(json.dumps(obj) + "\\n")
23
+ sys.stdout.flush()
24
+
25
+ def main():
26
+ try:
27
+ req = json.loads(sys.stdin.read())
28
+ except Exception as e:
29
+ emit({"event": "error", "message": f"bad request: {e}"})
30
+ sys.exit(2)
31
+
32
+ model = req.get("model")
33
+ messages = req.get("messages", [])
34
+ max_tokens = int(req.get("max_tokens") or 256)
35
+ temperature = float(req.get("temperature") or 0.7)
36
+ if not model:
37
+ emit({"event": "error", "message": "model required"})
38
+ sys.exit(2)
39
+
40
+ try:
41
+ import torch
42
+ from transformers import AutoTokenizer
43
+ from petals import AutoDistributedModelForCausalLM
44
+ except ImportError as e:
45
+ emit({"event": "error", "message": f"petals not installed: {e}. Install: pip install -U petals"})
46
+ sys.exit(2)
47
+
48
+ emit({"event": "loading", "model": model})
49
+ tokenizer = AutoTokenizer.from_pretrained(model)
50
+ model_obj = AutoDistributedModelForCausalLM.from_pretrained(model, torch_dtype=torch.float16)
51
+
52
+ # Compose a single prompt from messages using the tokenizer's chat template.
53
+ prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
54
+ input_ids = tokenizer.encode(prompt, return_tensors="pt")
55
+
56
+ emit({"event": "ready", "input_tokens": int(input_ids.shape[-1])})
57
+
58
+ # Petals streams via inference_session — emit one JSON line per generated token.
59
+ full_text = ""
60
+ t0 = time.time()
61
+ with model_obj.inference_session(max_length=int(input_ids.shape[-1]) + max_tokens) as session:
62
+ # Prefill
63
+ outputs = model_obj.generate(
64
+ input_ids,
65
+ max_new_tokens=1,
66
+ do_sample=temperature > 0,
67
+ temperature=max(temperature, 1e-3),
68
+ session=session,
69
+ )
70
+ new_id = int(outputs[0, -1])
71
+ token_text = tokenizer.decode([new_id], skip_special_tokens=True)
72
+ full_text += token_text
73
+ emit({"event": "token", "text": token_text})
74
+
75
+ for _ in range(max_tokens - 1):
76
+ outputs = model_obj.generate(
77
+ outputs,
78
+ max_new_tokens=1,
79
+ do_sample=temperature > 0,
80
+ temperature=max(temperature, 1e-3),
81
+ session=session,
82
+ )
83
+ new_id = int(outputs[0, -1])
84
+ if new_id == tokenizer.eos_token_id:
85
+ break
86
+ token_text = tokenizer.decode([new_id], skip_special_tokens=True)
87
+ full_text += token_text
88
+ emit({"event": "token", "text": token_text})
89
+
90
+ emit({"event": "done", "text": full_text, "elapsed_s": round(time.time() - t0, 2)})
91
+
92
+ if __name__ == "__main__":
93
+ try:
94
+ main()
95
+ except Exception as e:
96
+ emit({"event": "error", "message": str(e)})
97
+ sys.exit(1)
98
+ `;
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@infernetprotocol/cli",
3
- "version": "0.1.37",
3
+ "version": "0.1.39",
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",
@@ -30,11 +30,11 @@
30
30
  "ora": "^9.4.0",
31
31
  "react": "^17.0.2",
32
32
  "react-blessed": "^0.7.2",
33
- "@infernetprotocol/auth": "0.1.37",
34
- "@infernetprotocol/engine": "0.1.37",
35
- "@infernetprotocol/deploy-providers": "0.1.37",
36
- "@infernetprotocol/gpu": "0.1.37",
37
- "@infernetprotocol/config": "0.1.37"
33
+ "@infernetprotocol/config": "0.1.39",
34
+ "@infernetprotocol/auth": "0.1.39",
35
+ "@infernetprotocol/deploy-providers": "0.1.39",
36
+ "@infernetprotocol/engine": "0.1.39",
37
+ "@infernetprotocol/gpu": "0.1.39"
38
38
  },
39
39
  "repository": {
40
40
  "type": "git",