|
| 1 | +"""Local Model Door helpers. |
| 2 | +
|
| 3 | +This module probes local model runtime state and renders routing/profile plans. |
| 4 | +It never pulls weights, starts daemons, sends prompts, or performs inference. |
| 5 | +""" |
| 6 | + |
| 7 | +from __future__ import annotations |
| 8 | + |
| 9 | +import datetime as _dt |
| 10 | +import hashlib |
| 11 | +import json |
| 12 | +import shutil |
| 13 | +import subprocess |
| 14 | +import sys |
| 15 | +from pathlib import Path |
| 16 | +from typing import Any, Dict |
| 17 | + |
| 18 | + |
| 19 | +DEFAULT_PROFILE_REF = "urn:srcos:model-profile:local-llama32-1b" |
| 20 | +QUALITY_PROFILE_REF = "urn:srcos:model-profile:local-llama32-3b" |
| 21 | +DEFAULT_ROUTER_BINDING_REF = "urn:socioprophet:model-router-binding:demo-user-local-llama32" |
| 22 | + |
| 23 | +LOCAL_MODEL_PROFILES = { |
| 24 | + "local-llama32-1b": { |
| 25 | + "profileRef": DEFAULT_PROFILE_REF, |
| 26 | + "displayName": "Local Llama 3.2 1B Router", |
| 27 | + "runtime": "ollama", |
| 28 | + "model": "llama3.2:1b", |
| 29 | + "parameterClass": "1b", |
| 30 | + "roles": [ |
| 31 | + "router", |
| 32 | + "triage", |
| 33 | + "summarization", |
| 34 | + "rewrite", |
| 35 | + "office-assist", |
| 36 | + "agent-machine-assist", |
| 37 | + "offline-fallback", |
| 38 | + "privacy-first-chat", |
| 39 | + ], |
| 40 | + "policy": { |
| 41 | + "localOnlyDefault": True, |
| 42 | + "sendPromptOffDeviceDefault": False, |
| 43 | + "allowToolUse": False, |
| 44 | + "allowNetwork": False, |
| 45 | + "requiresExplicitPull": True, |
| 46 | + }, |
| 47 | + }, |
| 48 | + "local-llama32-3b": { |
| 49 | + "profileRef": QUALITY_PROFILE_REF, |
| 50 | + "displayName": "Local Llama 3.2 3B Quality Fallback", |
| 51 | + "runtime": "ollama", |
| 52 | + "model": "llama3.2:3b", |
| 53 | + "parameterClass": "3b", |
| 54 | + "roles": [ |
| 55 | + "summarization", |
| 56 | + "rewrite", |
| 57 | + "office-assist", |
| 58 | + "agent-machine-assist", |
| 59 | + "offline-fallback", |
| 60 | + "coding-assist", |
| 61 | + "privacy-first-chat", |
| 62 | + ], |
| 63 | + "policy": { |
| 64 | + "localOnlyDefault": True, |
| 65 | + "sendPromptOffDeviceDefault": False, |
| 66 | + "allowToolUse": False, |
| 67 | + "allowNetwork": False, |
| 68 | + "requiresExplicitPull": True, |
| 69 | + }, |
| 70 | + }, |
| 71 | +} |
| 72 | + |
| 73 | + |
| 74 | +def _print_json(payload: Dict[str, Any]) -> int: |
| 75 | + print(json.dumps(payload, indent=2, sort_keys=True)) |
| 76 | + return 0 |
| 77 | + |
| 78 | + |
| 79 | +def _ollama_path() -> str | None: |
| 80 | + return shutil.which("ollama") |
| 81 | + |
| 82 | + |
| 83 | +def _ollama_list() -> dict[str, Any]: |
| 84 | + path = _ollama_path() |
| 85 | + if not path: |
| 86 | + return {"available": False, "path": None, "models": [], "error": "ollama not found on PATH"} |
| 87 | + try: |
| 88 | + completed = subprocess.run( |
| 89 | + [path, "list"], |
| 90 | + check=False, |
| 91 | + capture_output=True, |
| 92 | + text=True, |
| 93 | + timeout=8, |
| 94 | + ) |
| 95 | + except Exception as exc: # pragma: no cover - defensive around host runtime |
| 96 | + return {"available": True, "path": path, "models": [], "error": str(exc)} |
| 97 | + |
| 98 | + models: list[str] = [] |
| 99 | + for line in completed.stdout.splitlines()[1:]: |
| 100 | + parts = line.split() |
| 101 | + if parts: |
| 102 | + models.append(parts[0]) |
| 103 | + return { |
| 104 | + "available": True, |
| 105 | + "path": path, |
| 106 | + "returnCode": completed.returncode, |
| 107 | + "models": models, |
| 108 | + "error": completed.stderr.strip() or None, |
| 109 | + } |
| 110 | + |
| 111 | + |
| 112 | +def _profile(profile_name: str) -> dict[str, Any]: |
| 113 | + try: |
| 114 | + return LOCAL_MODEL_PROFILES[profile_name] |
| 115 | + except KeyError: |
| 116 | + known = ", ".join(sorted(LOCAL_MODEL_PROFILES)) |
| 117 | + raise SystemExit(f"unknown local model profile: {profile_name}; known profiles: {known}") |
| 118 | + |
| 119 | + |
| 120 | +def _prompt_hash(text: str | None) -> str | None: |
| 121 | + if text is None: |
| 122 | + return None |
| 123 | + return "sha256:" + hashlib.sha256(text.encode("utf-8")).hexdigest() |
| 124 | + |
| 125 | + |
| 126 | +def doctor(args) -> int: |
| 127 | + """Inspect local model runtime availability without pulling or running models.""" |
| 128 | + runtime = _ollama_list() |
| 129 | + installed = set(runtime.get("models", [])) |
| 130 | + profiles = [] |
| 131 | + for profile in LOCAL_MODEL_PROFILES.values(): |
| 132 | + model = profile["model"] |
| 133 | + profiles.append( |
| 134 | + { |
| 135 | + "profileRef": profile["profileRef"], |
| 136 | + "model": model, |
| 137 | + "runtime": profile["runtime"], |
| 138 | + "installed": model in installed, |
| 139 | + "roles": profile["roles"], |
| 140 | + } |
| 141 | + ) |
| 142 | + return _print_json( |
| 143 | + { |
| 144 | + "type": "LocalModelDoctor", |
| 145 | + "capturedAt": _dt.datetime.now(_dt.timezone.utc).isoformat(), |
| 146 | + "runtime": runtime, |
| 147 | + "profiles": profiles, |
| 148 | + "policy": { |
| 149 | + "pullsModels": False, |
| 150 | + "startsServices": False, |
| 151 | + "runsInference": False, |
| 152 | + "promptEgressDefault": "deny", |
| 153 | + "promptEvidence": "hash-only", |
| 154 | + }, |
| 155 | + } |
| 156 | + ) |
| 157 | + |
| 158 | + |
| 159 | +def profiles(args) -> int: |
| 160 | + """List built-in local model profile refs consumed from sourceos-model-carry.""" |
| 161 | + return _print_json( |
| 162 | + { |
| 163 | + "type": "LocalModelProfiles", |
| 164 | + "profiles": LOCAL_MODEL_PROFILES, |
| 165 | + "sourceRepo": "SourceOS-Linux/sourceos-model-carry", |
| 166 | + } |
| 167 | + ) |
| 168 | + |
| 169 | + |
| 170 | +def plan(args) -> int: |
| 171 | + """Render a local model runtime plan without pulling or running models.""" |
| 172 | + profile = _profile(args.profile) |
| 173 | + runtime = _ollama_list() |
| 174 | + installed = profile["model"] in set(runtime.get("models", [])) |
| 175 | + return _print_json( |
| 176 | + { |
| 177 | + "type": "LocalModelPlan", |
| 178 | + "profile": profile, |
| 179 | + "runtime": runtime, |
| 180 | + "installed": installed, |
| 181 | + "wouldPull": False, |
| 182 | + "wouldStartService": False, |
| 183 | + "wouldRunInference": False, |
| 184 | + "explicitInstallCommand": f"ollama pull {profile['model']}", |
| 185 | + "explicitRunCommand": f"ollama run {profile['model']}", |
| 186 | + "policy": profile["policy"], |
| 187 | + } |
| 188 | + ) |
| 189 | + |
| 190 | + |
| 191 | +def route(args) -> int: |
| 192 | + """Render a hash-only route decision for a task class.""" |
| 193 | + runtime = _ollama_list() |
| 194 | + installed = set(runtime.get("models", [])) |
| 195 | + default = LOCAL_MODEL_PROFILES["local-llama32-1b"] |
| 196 | + quality = LOCAL_MODEL_PROFILES["local-llama32-3b"] |
| 197 | + has_default = default["model"] in installed |
| 198 | + has_quality = quality["model"] in installed |
| 199 | + |
| 200 | + target = "base-local" if has_default else "quality-local" if has_quality else "hosted-policy-required" |
| 201 | + if args.task_class in {"office-assist", "rewrite", "summarization"} and args.personalization_ref: |
| 202 | + target = "personal-local-policy-checked" |
| 203 | + |
| 204 | + return _print_json( |
| 205 | + { |
| 206 | + "type": "LocalModelRouteDecision", |
| 207 | + "capturedAt": _dt.datetime.now(_dt.timezone.utc).isoformat(), |
| 208 | + "taskClass": args.task_class, |
| 209 | + "target": target, |
| 210 | + "defaultLocalProfileRef": DEFAULT_PROFILE_REF, |
| 211 | + "qualityFallbackLocalProfileRef": QUALITY_PROFILE_REF, |
| 212 | + "routerBindingRef": args.router_binding_ref, |
| 213 | + "personalizationRef": args.personalization_ref, |
| 214 | + "promptHash": _prompt_hash(args.prompt) if args.prompt else None, |
| 215 | + "promptStored": False, |
| 216 | + "runtimeAvailable": runtime.get("available", False), |
| 217 | + "installedModels": sorted(installed), |
| 218 | + "requiresPolicyForHostedFallback": target == "hosted-policy-required", |
| 219 | + "evidence": { |
| 220 | + "emitRouteDecision": True, |
| 221 | + "emitRuntimeHealth": True, |
| 222 | + "emitGovernanceRefs": True, |
| 223 | + "promptHashOnly": True, |
| 224 | + }, |
| 225 | + } |
| 226 | + ) |
| 227 | + |
| 228 | + |
| 229 | +def evidence_inspect(args) -> int: |
| 230 | + path = Path(args.path) |
| 231 | + if not path.exists(): |
| 232 | + print(f"error: evidence file not found: {path}", file=sys.stderr) |
| 233 | + return 1 |
| 234 | + try: |
| 235 | + payload = json.loads(path.read_text()) |
| 236 | + except json.JSONDecodeError as exc: |
| 237 | + print(f"error: invalid JSON: {exc}", file=sys.stderr) |
| 238 | + return 1 |
| 239 | + return _print_json( |
| 240 | + { |
| 241 | + "path": str(path), |
| 242 | + "type": payload.get("type"), |
| 243 | + "taskClass": payload.get("taskClass"), |
| 244 | + "target": payload.get("target"), |
| 245 | + "promptStored": payload.get("promptStored"), |
| 246 | + "promptHashPresent": bool(payload.get("promptHash")), |
| 247 | + "routerBindingRef": payload.get("routerBindingRef"), |
| 248 | + } |
| 249 | + ) |
0 commit comments