@infernetprotocol/cli 0.1.70 → 0.1.71

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/setup.js CHANGED
@@ -47,6 +47,8 @@ Flags:
47
47
  --confirm, --yes Auto-confirm every prompt. Use for unattended runs.
48
48
  --model <name> Pre-select model to pull (e.g. qwen2.5:7b).
49
49
  --skip-pull Skip the model-pull step.
50
+ --no-abliterated Don't default to the abliterated (uncensored) model
51
+ when it fits; pick the largest stock Qwen instead.
50
52
  --backend <kind> Pin engine backend (vllm | ollama | mojo | stub). Default: auto (prefers vLLM, falls back to Ollama).
51
53
  --host <url> Override Ollama host. Default: http://localhost:11434.
52
54
  --port <n> P2P port to open in the firewall. Default: ${DEFAULT_P2P_PORT}.
@@ -69,15 +71,38 @@ skip the prompts.
69
71
  * is ~1.2× that for a Q4_K_M quant + KV cache. The model picker uses
70
72
  * this to decide whether a model fits the host.
71
73
  */
74
+ // huihui-ai's abliteration of Qwen3.8 27B (Q4_K_M on Ollama; the same weights
75
+ // are on HF as huihui-ai/Huihui-Qwen3.8-27B-abliterated for vLLM nodes).
76
+ export const ABLITERATED_DEFAULT = "huihui_ai/Qwen3.8-abliterated:27b";
77
+
72
78
  const MODEL_SUGGESTIONS = [
73
79
  { name: "qwen2.5:0.5b", size_gb: 0.4, size: "≈400 MB", note: "smoke test, runs on CPU" },
74
80
  { name: "qwen2.5:3b", size_gb: 2.0, size: "≈2 GB", note: "fits 6 GB+ GPU / 8 GB+ RAM" },
75
81
  { name: "qwen2.5:7b", size_gb: 4.4, size: "≈4.4 GB", note: "fits 8 GB+ GPU / 12 GB+ RAM" },
76
82
  { name: "qwen2.5:14b", size_gb: 9.0, size: "≈9 GB", note: "fits 16 GB+ GPU / 24 GB+ RAM" },
77
83
  { name: "qwen2.5:32b", size_gb: 20.0, size: "≈20 GB", note: "fits 24 GB+ GPU / 40 GB+ RAM" },
78
- { name: "qwen2.5:72b", size_gb: 40.0, size: "≈40 GB", note: "fits 48 GB+ GPU / 80 GB+ RAM" }
84
+ { name: "qwen2.5:72b", size_gb: 40.0, size: "≈40 GB", note: "fits 48 GB+ GPU / 80 GB+ RAM" },
85
+ { name: ABLITERATED_DEFAULT, size_gb: 18.0, size: "≈18 GB", note: "abliterated (uncensored), fits 24 GB+ GPU", abliterated: true }
79
86
  ];
80
87
 
88
+ /**
89
+ * Default pick for a self-hosted node. CPU-only hosts get the smallest model.
90
+ * GPU hosts get the abliterated model when it fits (unless the operator opts
91
+ * out), otherwise the largest stock model that fits.
92
+ */
93
+ export function pickDefaultSuggestion(evaluated, { isCpuOnly, abliterated = true } = {}) {
94
+ if (isCpuOnly) return MODEL_SUGGESTIONS[0].name;
95
+ const fitting = evaluated.filter((m) => m.fits);
96
+ if (abliterated) {
97
+ const abl = fitting.find((m) => m.abliterated);
98
+ if (abl) return abl.name;
99
+ }
100
+ const stock = fitting.filter((m) => !m.abliterated);
101
+ return stock.length > 0 ? stock[stock.length - 1].name : MODEL_SUGGESTIONS[0].name;
102
+ }
103
+
104
+ export { MODEL_SUGGESTIONS };
105
+
81
106
  // modelFits wraps checkFits: maps ok→fits and passes through unknown sizes.
82
107
  export function modelFits({ size_gb, vram_gb, ram_gb }) {
83
108
  if (!Number.isFinite(size_gb) || size_gb <= 0) {
@@ -392,7 +417,9 @@ async function verifyModelPulled(host, name) {
392
417
  clearTimeout(t);
393
418
  if (!res.ok) return false;
394
419
  const body = await res.json();
395
- const have = (body.models ?? []).some((m) => (m.name ?? m.model) === name);
420
+ // Ollama may report namespaced tags in lower case.
421
+ const want = name.toLowerCase();
422
+ const have = (body.models ?? []).some((m) => String(m.name ?? m.model).toLowerCase() === want);
396
423
  return have;
397
424
  } catch {
398
425
  return false;
@@ -408,15 +435,11 @@ async function chooseModel(installed, opts) {
408
435
  const f = checkFits({ size_gb: m.size_gb, vram_gb: vramGb, ram_gb: ramGb });
409
436
  return { ...m, fits: f.ok, mode: f.mode, ceiling_gb: f.ceiling_gb };
410
437
  });
411
- const fittingOnly = evaluated.filter((m) => m.fits);
412
438
  const isCpuOnly = !(Number.isFinite(vramGb) && vramGb > 0);
413
439
  // CPU-only: default to smallest model — large models are impractically
414
- // slow on CPU and will lock the machine. GPU: largest fitting model.
415
- const defaultModel = isCpuOnly
416
- ? MODEL_SUGGESTIONS[0].name // qwen2.5:0.5b
417
- : fittingOnly.length > 0
418
- ? fittingOnly[fittingOnly.length - 1].name
419
- : MODEL_SUGGESTIONS[0].name;
440
+ // slow on CPU and will lock the machine. GPU: the abliterated model if it
441
+ // fits, else the largest fitting stock model.
442
+ const defaultModel = pickDefaultSuggestion(evaluated, { isCpuOnly, abliterated: opts.abliterated !== false });
420
443
 
421
444
  if (opts.preselected) {
422
445
  const m = evaluated.find((x) => x.name === opts.preselected);
@@ -441,7 +464,7 @@ async function chooseModel(installed, opts) {
441
464
  ? ""
442
465
  : ` [✗ won't fit on this host — ${m.mode} ceiling ≈ ${m.ceiling_gb} GB]`;
443
466
  process.stdout.write(
444
- ` ${i + 1}) ${m.name.padEnd(16)} ${m.size.padEnd(10)} ${m.note}${installedFlag}${fitFlag}\n`
467
+ ` ${i + 1}) ${m.name.padEnd(34)} ${m.size.padEnd(10)} ${m.note}${installedFlag}${fitFlag}\n`
445
468
  );
446
469
  });
447
470
  process.stdout.write(` ${MODEL_SUGGESTIONS.length + 1}) other (enter manually)\n`);
@@ -679,6 +702,7 @@ export default async function setup(args) {
679
702
  chosenModel = await chooseModel(installedNames, {
680
703
  yes,
681
704
  preselected: preselectedModel,
705
+ abliterated: !args.has("no-abliterated"),
682
706
  capacity: { vramGb, ramGb }
683
707
  });
684
708
  if (!chosenModel) {
@@ -37,12 +37,14 @@ How it works:
37
37
 
38
38
  Curated models (HuggingFace / vLLM):
39
39
  ≥ 8 GB VRAM NousResearch/Hermes-3-Llama-3.1-8B (Apache 2.0, ~5 GB @ 4-bit)
40
- ≥ 40 GB VRAM cognitivecomputations/dolphin-2.9.4-llama3.1-70b
40
+ ≥ 40 GB VRAM munekazu/Huihui-Qwen3.8-27B-abliterated-FP8
41
+ ≥ 256 GB GLM-5.3 Flash / GLM-5.3 / DeepSeek V4.1 Flash abliterated
42
+ (multi-GPU; see --list)
41
43
 
42
44
  Curated models (Ollama):
43
45
  ≥ 4 GB VRAM dolphin3:2b
44
46
  ≥ 6 GB VRAM dolphin3:8b (recommended)
45
- ≥ 30 GB VRAM dolphin-mixtral:8x7b
47
+ ≥ 24 GB VRAM huihui_ai/Qwen3.8-abliterated:27b
46
48
  `;
47
49
 
48
50
  // HuggingFace models, ordered by VRAM requirement ascending
@@ -68,6 +70,38 @@ const HF_MODELS = [
68
70
  minVramGb: 40,
69
71
  description: 'Dolphin 2.9.4 70B (Eric Hartford) — uncensored, Llama 3.1 base',
70
72
  },
73
+ {
74
+ // huihui-ai abliteration of Qwen3.8 27B, FP8 (~31 GB). Newest and
75
+ // strongest single-GPU pick, so it outranks the 70B when both fit.
76
+ repoId: 'munekazu/Huihui-Qwen3.8-27B-abliterated-FP8',
77
+ sizeGb: 55.6,
78
+ sizeGb4bit: 18.0,
79
+ minVramGb: 40,
80
+ description: 'Qwen3.8 27B abliterated (huihui-ai) — uncensored, FP8 via vLLM, Apache 2.0',
81
+ },
82
+ {
83
+ // The open-weight abliterated models allmodels.io routes to. MoE
84
+ // flagships: multi-GPU only. NVFP4 builds need Blackwell (B200/GB200).
85
+ repoId: 'edp1096/Huihui-GLM-5.3-Flash-abliterated-NVFP4',
86
+ sizeGb: 204.4,
87
+ sizeGb4bit: 204.4,
88
+ minVramGb: 256,
89
+ description: 'GLM-5.3 Flash 321B abliterated (huihui-ai) — NVFP4, needs Blackwell GPUs (~256 GB VRAM)',
90
+ },
91
+ {
92
+ repoId: 'dealignai/GLM-5.3-ABLITERATED-NVFP4',
93
+ sizeGb: 464.8,
94
+ sizeGb4bit: 464.8,
95
+ minVramGb: 512,
96
+ description: 'GLM-5.3 753B abliterated — NVFP4, needs Blackwell GPUs (~512 GB VRAM)',
97
+ },
98
+ {
99
+ repoId: 'mrfakename/DeepSeek-V4.1-Flash-1017-abliterated',
100
+ sizeGb: 510.3,
101
+ sizeGb4bit: 510.3,
102
+ minVramGb: 640,
103
+ description: 'DeepSeek V4.1 Flash 763B abliterated — FP8, needs 8x80 GB GPUs',
104
+ },
71
105
  {
72
106
  // 754B MoE, FP8 weights (~755 GB) — flagship uncensored model. Only
73
107
  // auto-selected on a cluster that can actually hold it (minVramGb set
@@ -86,6 +120,8 @@ const OLLAMA_MODELS = [
86
120
  { name: 'dolphin3:2b', sizeGb: 2.2, minVramGb: 4, description: 'Dolphin 3 2B — lightweight uncensored' },
87
121
  { name: 'dolphin3:8b', sizeGb: 4.9, minVramGb: 6, description: 'Dolphin 3 8B — recommended uncensored (Llama 3.1)' },
88
122
  { name: 'dolphin-mixtral:8x7b', sizeGb: 26.1, minVramGb: 30, description: 'Dolphin Mixtral 8x7B — high-capacity uncensored' },
123
+ // Last so it wins whenever it fits: newer and stronger than Mixtral.
124
+ { name: 'huihui_ai/Qwen3.8-abliterated:27b', sizeGb: 18.0, minVramGb: 24, description: 'Qwen3.8 27B abliterated (huihui-ai) — recommended on 24 GB+' },
89
125
  ];
90
126
 
91
127
  function pickHfModel(vramGb) {
@@ -52,7 +52,13 @@ export const CATALOG = [
52
52
  // ── MoE (high VRAM, vLLM only) ──
53
53
  { id: "mixtral-8x7b", name: "Mixtral 8x7B", paramsB: 47, vramMin: 30, ramMin: 64, pullName: "dolphin-mixtral:8x7b", backend: "ollama", quality: 9, speed: 5, useCases: ["chat","coding","uncensored"], uncensored: true },
54
54
 
55
+ // ── abliterated (the open-weight models allmodels.io routes as "abliterated") ──
56
+ { id: "qwen3.8-27b-abliterated", name: "Qwen3.8 27B abliterated (uncensored)", paramsB: 27.8, vramMin: 24, ramMin: 32, pullName: "huihui_ai/Qwen3.8-abliterated:27b", backend: "ollama", quality: 10, speed: 5, useCases: ["chat","uncensored","coding","agents","study"], uncensored: true, abliterated: true, notes: "huihui-ai abliteration; default for 24 GB+ self-hosted nodes." },
57
+
55
58
  // ── flagship / cluster-tier (multi-GPU or cloud; won't fit a single card) ──
59
+ { id: "glm-5.3-flash-abliterated", name: "GLM-5.3 Flash 321B NVFP4 (abliterated)", paramsB: 321, vramMin: 256, ramMin: 256, pullName: "hf:edp1096/Huihui-GLM-5.3-Flash-abliterated-NVFP4", backend: "vllm", quality: 10, speed: 3, useCases: ["chat","uncensored","coding","agents"], uncensored: true, abliterated: true, notes: "NVFP4 needs Blackwell GPUs." },
60
+ { id: "glm-5.3-abliterated", name: "GLM-5.3 753B NVFP4 (abliterated)", paramsB: 753, vramMin: 512, ramMin: 512, pullName: "hf:dealignai/GLM-5.3-ABLITERATED-NVFP4", backend: "vllm", quality: 10, speed: 2, useCases: ["chat","uncensored","coding","agents"], uncensored: true, abliterated: true, notes: "NVFP4 needs Blackwell GPUs." },
61
+ { id: "deepseek-v4.1-flash-abliterated", name: "DeepSeek V4.1 Flash 763B (abliterated)", paramsB: 763, vramMin: 640, ramMin: 640, pullName: "hf:mrfakename/DeepSeek-V4.1-Flash-1017-abliterated", backend: "vllm", quality: 10, speed: 2, useCases: ["chat","uncensored","coding","agents","reasoning"], uncensored: true, abliterated: true },
56
62
  { id: "glm-5.2-uncensored", name: "GLM-5.2 754B FP8 (uncensored)", paramsB: 754, vramMin: 768, ramMin: 768, pullName: "hf:zandenAI/GLM-5.2-FP8-Uncensored", backend: "vllm", quality: 10, speed: 2, useCases: ["chat","uncensored","cybersecurity","agents"], uncensored: true, gated: true }
57
63
  ];
58
64
 
@@ -78,6 +84,8 @@ function scoreModel(m, { vramGb, ramGb, useCase }) {
78
84
  // Quality bias for use-case match
79
85
  score += m.quality * 4;
80
86
  if (useCase && m.useCases?.includes(useCase)) score += 25;
87
+ // Self-hosted default: prefer an abliterated model whenever it fits.
88
+ if (m.abliterated && (m.vramMin === 0 || vramGb >= m.vramMin)) score += 20;
81
89
 
82
90
  // CPU-only systems prefer speed
83
91
  if (vramGb === 0) score += m.speed * 3 - m.paramsB * 2;
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@infernetprotocol/cli",
3
- "version": "0.1.70",
3
+ "version": "0.1.71",
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",
@@ -31,13 +31,13 @@
31
31
  "ora": "^9.4.1",
32
32
  "react": "^17.0.2",
33
33
  "react-blessed": "^0.7.2",
34
- "@infernetprotocol/config": "0.1.70",
35
- "@infernetprotocol/auth": "0.1.70",
36
- "@infernetprotocol/deploy-providers": "0.1.70",
37
- "@infernetprotocol/engine": "0.1.70",
38
- "@infernetprotocol/liquid-agent": "0.1.70",
39
- "@infernetprotocol/rpc-adapter": "0.1.70",
40
- "@infernetprotocol/gpu": "0.1.70"
34
+ "@infernetprotocol/auth": "0.1.71",
35
+ "@infernetprotocol/config": "0.1.71",
36
+ "@infernetprotocol/deploy-providers": "0.1.71",
37
+ "@infernetprotocol/engine": "0.1.71",
38
+ "@infernetprotocol/gpu": "0.1.71",
39
+ "@infernetprotocol/liquid-agent": "0.1.71",
40
+ "@infernetprotocol/rpc-adapter": "0.1.71"
41
41
  },
42
42
  "repository": {
43
43
  "type": "git",