File size: 10,483 Bytes
6343479
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
"""Central Autonomous Resource Manager for FlyBrain V8/V9.

Implements Sections 11–22:
- Dynamic empirical system probing
- Subsystem budget allocations
- Model lifecycle & admission control
- Priority-driven onload/offload (P0 to P5)
- Hysteresis pressure tracking (GREEN to CRITICAL)
- Bounded OOM recovery & quarantine
"""
import os
import gc
import time
import logging
import threading
from enum import Enum
from typing import Dict, Any, Optional, Callable, List

from src.runtime.resource_manager.prober import probe_system, probe_memory
from src.runtime.resource_manager.profile import derive_profile, compute_budgets, ProfileTier, SubsystemBudgets
from src.runtime.resource_manager.pressure import PressureEvaluator, PressureState

logger = logging.getLogger("FlyBrain.ResourceManager")


class ModelLifecycle(str, Enum):
    DISCOVERED = "DISCOVERED"
    INDEXED = "INDEXED"
    AVAILABLE = "AVAILABLE"
    LOADING = "LOADING"
    LOADED = "LOADED"
    WARM = "WARM"
    BUSY = "BUSY"
    IDLE = "IDLE"
    DRAINING = "DRAINING"
    OFFLOADING = "OFFLOADING"
    OFFLOADED = "OFFLOADED"
    FAILED = "FAILED"
    QUARANTINED = "QUARANTINED"


class ModelPriority(int, Enum):
    P0_CORE = 0         # Brain controller, life simulation (NEVER EVICTED)
    P1_INTERACTIVE = 1  # Navigation, direct dialogue, active VLM
    P2_SIMULATION = 2   # Embeddings, spatial memory search
    P3_RESEARCH = 3     # Image generation, counterfactual replay
    P4_BACKGROUND = 4   # Dream synthesis, image->3D
    P5_SPECULATIVE = 5  # Pre-baked assets, speculative variation (FIRST EVICTED)


class ResourceManager:
    _instance = None
    _lock = threading.RLock()

    def __new__(cls, *args, **kwargs):
        with cls._lock:
            if cls._instance is None:
                cls._instance = super(ResourceManager, cls).__new__(cls)
                cls._instance._initialized = False
            return cls._instance

    def __init__(self, override_tier: Optional[ProfileTier] = None):
        if self._initialized:
            return
        self._lock = threading.RLock()
        self.probe = probe_system()
        self.tier = override_tier or derive_profile(self.probe)
        self.budgets = compute_budgets(self.probe, self.tier)
        self.pressure_eval = PressureEvaluator()
        self.models: Dict[str, Dict[str, Any]] = {}
        self.oom_retries: Dict[str, int] = {}
        self.events: List[Dict[str, Any]] = []
        self._initialized = True
        self.log_event("INITIALIZED", f"Tier={self.tier.value}, SafeRAM={self.budgets.safe_ram_target_mb}MB")

    def log_event(self, kind: str, detail: str) -> None:
        event = {"ts": time.time(), "kind": kind, "detail": detail}
        self.events.append(event)
        if len(self.events) > 500:
            self.events.pop(0)
        logger.info(f"[{kind}] {detail}")

    def refresh_telemetry(self) -> Dict[str, Any]:
        """Refreshes live memory and updates pressure state."""
        mem = probe_memory()
        state = self.pressure_eval.evaluate(mem["ram_used_percent"])
        
        if state in (PressureState.RED, PressureState.CRITICAL):
            self._handle_high_pressure(state)
            
        return {
            "tier": self.tier.value,
            "pressure": self.pressure_eval.status(),
            "memory": mem,
            "budgets": self.budgets.to_dict(),
            "loaded_models": [k for k, v in self.models.items() if v["state"] in (ModelLifecycle.LOADED, ModelLifecycle.WARM, ModelLifecycle.BUSY)]
        }

    def register_model(self, model_id: str, priority: ModelPriority,
                       ram_estimate_mb: float, vram_estimate_mb: float,
                       loader: Callable[[], Any], unloader: Callable[[Any], None]) -> None:
        """Registers a model into managed lifecycle."""
        with self._lock:
            self.models[model_id] = {
                "id": model_id,
                "priority": priority,
                "ram_mb": ram_estimate_mb,
                "vram_mb": vram_estimate_mb,
                "state": ModelLifecycle.AVAILABLE,
                "instance": None,
                "loader": loader,
                "unloader": unloader,
                "last_used": 0.0,
                "use_count": 0,
                "quarantined": False
            }

    def request_model(self, model_id: str) -> Optional[Any]:
        """Admission-controlled model acquisition."""
        with self._lock:
            spec = self.models.get(model_id)
            if not spec:
                raise ValueError(f"Unknown model {model_id}")

            if spec["state"] == ModelLifecycle.QUARANTINED:
                raise RuntimeError(f"Model {model_id} is quarantined due to repeated failures.")

            if spec["state"] in (ModelLifecycle.LOADED, ModelLifecycle.WARM):
                spec["last_used"] = time.time()
                spec["use_count"] += 1
                return spec["instance"]

            # Admission check
            if not self._check_admission(spec):
                self._evict_for_admission(spec["ram_mb"])
                if not self._check_admission(spec):
                    raise MemoryError(f"Admission denied for {model_id}: Insufficient resource budget.")

            # Load model with OOM guard
            return self._load_model_safe(model_id)

    def _check_admission(self, spec: Dict[str, Any]) -> bool:
        mem = probe_memory()
        available_mb = mem["ram_available_mb"] if "ram_available_mb" in mem else (mem["ram_available_bytes"] / (1024 ** 2))
        safety_margin_mb = self.budgets.os_reserve_mb * 0.5
        return (available_mb - spec["ram_mb"]) >= safety_margin_mb

    def _load_model_safe(self, model_id: str) -> Any:
        spec = self.models[model_id]
        spec["state"] = ModelLifecycle.LOADING
        self.log_event("MODEL_LOADING", f"Loading {model_id} (Est: {spec['ram_mb']}MB)")
        
        try:
            instance = spec["loader"]()
            spec["instance"] = instance
            spec["state"] = ModelLifecycle.LOADED
            spec["last_used"] = time.time()
            spec["use_count"] += 1
            self.log_event("MODEL_LOADED", f"Successfully loaded {model_id}")
            return instance
        except Exception as e:
            self.log_event("MODEL_FAILED", f"Error loading {model_id}: {e}")
            return self._recover_oom(model_id, e)

    def release_model(self, model_id: str) -> None:
        """Unloads an active model releasing host and device memory."""
        with self._lock:
            spec = self.models.get(model_id)
            if not spec or spec["state"] not in (ModelLifecycle.LOADED, ModelLifecycle.WARM, ModelLifecycle.BUSY):
                return
            
            spec["state"] = ModelLifecycle.OFFLOADING
            self.log_event("MODEL_OFFLOADING", f"Evicting {model_id}")
            try:
                if spec["instance"] is not None and spec["unloader"]:
                    spec["unloader"](spec["instance"])
            except Exception as e:
                logger.warning(f"Error during unload of {model_id}: {e}")
            finally:
                spec["instance"] = None
                spec["state"] = ModelLifecycle.OFFLOADED
                gc.collect()
                self.log_event("MODEL_OFFLOADED", f"Freed memory for {model_id}")

    def _evict_for_admission(self, required_mb: float) -> None:
        """Evicts lower priority models (P5 down to P1) until required_mb is freed."""
        candidates = [m for m in self.models.values() if m["state"] in (ModelLifecycle.LOADED, ModelLifecycle.WARM)]
        # Sort by priority DESC (P5 first, P0 never), then by last_used ASC
        candidates.sort(key=lambda x: (x["priority"].value, -x["last_used"]), reverse=True)

        for cand in candidates:
            if cand["priority"] == ModelPriority.P0_CORE:
                continue # Never evict P0
            self.release_model(cand["id"])
            mem = probe_memory()
            avail_mb = mem["ram_available_bytes"] / (1024 ** 2)
            if avail_mb >= required_mb + self.budgets.os_reserve_mb * 0.5:
                break

    def _handle_high_pressure(self, state: PressureState) -> None:
        """Automated response to resource pressure."""
        self.log_event("PRESSURE_RESPONSE", f"Triggered response for {state.value}")
        # Evict P5 speculative and P4 background models
        for m in list(self.models.values()):
            if m["state"] in (ModelLifecycle.LOADED, ModelLifecycle.WARM):
                if state == PressureState.CRITICAL and m["priority"].value >= ModelPriority.P2_SIMULATION.value:
                    self.release_model(m["id"])
                elif state == PressureState.RED and m["priority"].value >= ModelPriority.P3_RESEARCH.value:
                    self.release_model(m["id"])
                elif state == PressureState.ORANGE and m["priority"].value >= ModelPriority.P4_BACKGROUND.value:
                    self.release_model(m["id"])

    def _recover_oom(self, model_id: str, error: Exception) -> Optional[Any]:
        """Bounded OOM recovery policy (Section 20)."""
        retries = self.oom_retries.get(model_id, 0)
        self.log_event("OOM_RECOVERY_ATTEMPT", f"Model {model_id} retry #{retries + 1}")
        
        if retries >= 2:
            spec = self.models[model_id]
            spec["state"] = ModelLifecycle.QUARANTINED
            self.log_event("MODEL_QUARANTINED", f"Quarantined {model_id} after {retries} failed attempts")
            raise RuntimeError(f"OOM recovery exhausted for {model_id}: {error}")

        self.oom_retries[model_id] = retries + 1
        gc.collect()
        
        # Evict all non-essential models
        for m in list(self.models.values()):
            if m["id"] != model_id and m["priority"].value >= ModelPriority.P2_SIMULATION.value:
                self.release_model(m["id"])
                
        # Re-attempt load
        return self._load_model_safe(model_id)

    def get_status(self) -> Dict[str, Any]:
        with self._lock:
            return {
                "tier": self.tier.value,
                "pressure": self.pressure_eval.status(),
                "budgets": self.budgets.to_dict(),
                "registered_models": {k: {"state": v["state"].value, "priority": v["priority"].name, "ram_mb": v["ram_mb"]} for k, v in self.models.items()},
                "recent_events": self.events[-15:]
            }