Spaces:
Running
Running
File size: 19,319 Bytes
e59bd9d | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 | """Adaptive Optimization Engine (mission §2.E, §7, §8).
Lifecycle of every change:
PROPOSED -> (validated? / user-confirmed?) -> APPLY-ROUTE -> APPLIED
|-> QUEUED_AT_BOUNDARY
|-> NEEDS_SDK
|-> NEEDS_RESTART
any applied/queued change -> ROLLBACK (restores the old value)
Safety rules (mission §7 levels + §21 rules 9/10/13/15):
- level 3 (LR / optimizer / loss / scheduler / EMA / precision / sampling /
augmentation / validation protocol) is NEVER auto-applied; every single
application requires the user's explicit confirmation, and we still only
route it through the SDK boundary or a restart instruction.
- level 2 requires either a passed controlled validation (measured A/B via
the SDK) or a recorded user override ("user_override_no_validation").
- level 1 may auto-apply only when config optimizer.auto_apply_level1=true
AND an SDK channel is attached; otherwise it is a manual proposal.
- restart-required parameters are NEVER claimed as "applied": the ledger
records status needs_restart plus the exact instruction.
- expected impact is always labeled "Estimate Only"; confidence is
"Measured" only when the proposal was derived from measured evidence.
Every state transition is journaled (in-memory ledger + SQLite param_changes
via the storage callback) so the whole decision history is auditable.
"""
from __future__ import annotations
import itertools
import time
import uuid
from dataclasses import dataclass, field
from typing import Callable, Dict, List, Optional
from .parameters import PARAM_SPECS, ParamSpec, ChangeClass, get_spec
APPLY_STATUSES = (
"applied", # SDK acked the change (hot-reload or boundary)
"queued_at_boundary", # SDK accepted, will ack at next batch/update/epoch edge
"needs_sdk", # no SDK channel: manual instruction recorded
"needs_restart", # restart-required class: never applied live
"failed", # SDK refused or timed out (nothing changed)
)
@dataclass
class Validation:
method: str # "sdk_ab_window" | "user_override" | "none"
before_metric: Optional[float] = None
after_metric: Optional[float] = None
metric_name: Optional[str] = None
verdict: str = "not_run" # passed | failed | not_run | user_override
detail: str = ""
def as_dict(self) -> dict:
return {
"method": self.method,
"metric_name": self.metric_name,
"before": self.before_metric,
"after": self.after_metric,
"verdict": self.verdict,
"detail": self.detail,
}
@dataclass
class Proposal:
id: str
param: str
current: object
proposed: object
level: int
change_class: str
reason: str
evidence: Dict[str, object]
confidence: str # Measured | Estimated
ts: float = field(default_factory=time.time)
status: str = "proposed" # proposed | approved | rejected |
# applied | queued_at_boundary |
# needs_sdk | needs_restart | failed |
# rolled_back
validation: Validation = field(default_factory=lambda: Validation(method="none"))
user_confirmed: bool = False
expected_impact: str = "Estimate Only"
def as_dict(self) -> dict:
return {
"id": self.id, "param": self.param,
"current": self.current, "proposed": self.proposed,
"level": self.level, "change_class": self.change_class,
"reason": self.reason, "evidence": self.evidence,
"confidence": self.confidence, "ts": self.ts,
"status": self.status, "validation": self.validation.as_dict(),
"user_confirmed": self.user_confirmed,
"expected_impact": self.expected_impact,
}
@dataclass
class ChangeRecord:
id: str
proposal_id: str
param: str
old_value: object
new_value: object
level: int
reason: str
status: str # same apply statuses + rolled_back
applied_ts: Optional[float] = None
result: str = ""
rollback_ts: Optional[float] = None
rollback_result: str = ""
def as_dict(self) -> dict:
return {
"id": self.id, "proposal_id": self.proposal_id,
"param": self.param, "old_value": self.old_value,
"new_value": self.new_value, "level": self.level,
"reason": self.reason, "status": self.status,
"applied_ts": self.applied_ts, "result": self.result,
"rollback_ts": self.rollback_ts,
"rollback_result": self.rollback_result,
}
# callback the service injects: (param, value, boundary: bool, extra: dict)
# -> {"status": one of APPLY_STATUSES, "detail": str}
CommandSender = Callable[[str, object, bool, dict], Dict[str, object]]
# callback to persist change records: (ChangeRecord.as_dict()) -> None
LedgerWriter = Callable[[dict], None]
class OptimizerEngine:
def __init__(self, config: Optional[dict] = None,
command_sender: Optional[CommandSender] = None,
ledger_writer: Optional[LedgerWriter] = None):
cfg = (config or {}).get("optimizer", {})
self.enabled = bool(cfg.get("enabled", True))
self.auto_apply_level1 = bool(cfg.get("auto_apply_level1", False))
self.cooldown_s = float(cfg.get("cooldown_s", 900))
self.max_proposals = int(cfg.get("max_proposals", 8))
self._sender = command_sender
self._ledger_writer = ledger_writer
self.proposals: Dict[str, Proposal] = {}
self.ledger: List[ChangeRecord] = []
self._ids = itertools.count(1)
self._last_proposed: Dict[str, float] = {}
# ------------------------------------------------------------------ propose
def _new_id(self, prefix: str) -> str:
return f"{prefix}-{next(self._ids)}-{uuid.uuid4().hex[:6]}"
def propose(self, param: str, current, proposed, reason: str,
evidence: Dict[str, object], confidence: str = "Measured") -> Optional[Proposal]:
if not self.enabled:
return None
try:
spec: ParamSpec = get_spec(param)
except KeyError:
return None # unregistered parameters are never proposed
now = time.time()
if now - self._last_proposed.get(param, 0.0) < self.cooldown_s:
return None
# dedup: same param + same target already on the board
for p in self.proposals.values():
if p.param == param and p.proposed == proposed and p.status in (
"proposed", "approved", "queued_at_boundary", "needs_sdk",
"needs_restart"):
return None
if len([p for p in self.proposals.values()
if p.status == "proposed"]) >= self.max_proposals:
return None
self._last_proposed[param] = now
prop = Proposal(
id=self._new_id("prop"), param=param, current=current,
proposed=proposed, level=int(spec.level),
change_class=spec.change_class.value, reason=reason,
evidence=evidence, confidence=confidence,
)
self.proposals[prop.id] = prop
return prop
def evaluate(self, facts: Dict[str, object],
bottleneck: Optional[dict]) -> List[Proposal]:
"""Derive proposals from measured facts + the current bottleneck
report. Pure decision logic; the sender applies nothing here."""
made: List[Proposal] = []
if not self.enabled:
return made
cat = (bottleneck or {}).get("category", "")
dwr = facts.get("data_wait_ratio_mean")
workers = facts.get("num_workers")
prefetch = facts.get("prefetch_factor")
pin = facts.get("pin_memory")
persistent = facts.get("persistent_workers")
gpu_idle_frac = facts.get("gpu_idle_fraction")
vram_used_pct = facts.get("vram_used_pct_mean")
oom_events = facts.get("oom_events")
micro = facts.get("micro_batch_size")
accum = facts.get("gradient_accumulation")
if "DataLoader Bound" in cat and dwr is not None and dwr >= 0.5:
if workers is not None and workers < 4:
made.append(self.propose(
"dataloader.num_workers", workers,
max(2, min(8, workers * 2 if workers else 4)),
"Data-wait ratio {:.2f} >= 0.5: GPUs are starving while "
"batches are being prepared".format(round(dwr, 3)),
{"data_wait_ratio": round(dwr, 3),
"gpu_idle_fraction": gpu_idle_frac,
"current_workers": workers},
confidence="Measured") or None)
if prefetch is not None and prefetch < 2:
made.append(self.propose(
"dataloader.prefetch_factor", prefetch, 2,
"Prefetch depth 1 leaves no overlap between batch "
"preparation and compute",
{"current_prefetch_factor": prefetch,
"data_wait_ratio": round(dwr or 0.0, 3)},
confidence="Measured") or None)
if pin is False:
made.append(self.propose(
"dataloader.pin_memory", False, True,
"Unpinned memory forces synchronous H2D copies",
{"current_pin_memory": False,
"note": "level 1 data-path change; no math effect"},
confidence="Measured") or None)
if persistent is False and (workers or 0) > 0:
made.append(self.propose(
"dataloader.persistent_workers", False, True,
"Windows spawn re-imports the dataset module every epoch; "
"persistent workers pay that cost once",
{"current_persistent_workers": False, "workers": workers},
confidence="Measured") or None)
if "VRAM Capacity Bound" in cat and vram_used_pct is not None:
if micro is not None and micro > 1 and (oom_events or 0) > 0:
made.append(self.propose(
"training.micro_batch_size", micro, micro // 2,
"VRAM used {:.1f}% with OOM events; halve micro-batch and "
"double accumulation to keep effective batch constant "
"(requires validation)".format(round(vram_used_pct, 1)),
{"vram_used_pct": round(vram_used_pct, 1),
"oom_events": oom_events,
"effective_batch_preserved": True},
confidence="Measured") or None)
return [p for p in made if p]
# ------------------------------------------------------------------ approve
def approve(self, proposal_id: str) -> Optional[Proposal]:
p = self.proposals.get(proposal_id)
if p and p.status == "proposed":
p.status = "approved"
p.user_confirmed = True
return p
def reject(self, proposal_id: str) -> Optional[Proposal]:
p = self.proposals.get(proposal_id)
if p and p.status in ("proposed", "approved"):
p.status = "rejected"
return p
def record_validation(self, proposal_id: str, validation: Validation) -> Optional[Proposal]:
p = self.proposals.get(proposal_id)
if p:
p.validation = validation
return p
# -------------------------------------------------------------------- apply
def apply(self, proposal_id: str, user_confirmed: bool = False,
override_validation: bool = False) -> dict:
p = self.proposals.get(proposal_id)
if p is None:
return {"ok": False, "error": "unknown proposal"}
if p.status not in ("proposed", "approved", "needs_sdk"):
return {"ok": False, "error": f"proposal status is {p.status!r}; "
"only proposed/approved can be applied"}
spec = PARAM_SPECS[p.param]
# --- gate 1: level 3 always requires this explicit confirmation ---
if spec.level == 3 and not (user_confirmed or p.user_confirmed):
return {"ok": False, "error": "level-3 parameter: explicit user "
"confirmation required for EVERY apply"}
# --- gate 2: level 2 requires a passed validation or recorded override
if spec.validation_required:
verdict = p.validation.verdict
if verdict != "passed":
if override_validation and (user_confirmed or p.user_confirmed):
p.validation = Validation(
method="user_override", verdict="user_override",
detail="user explicitly applied without validation")
else:
return {"ok": False, "error": "level-2 parameter needs a "
"passed validation (or an "
"explicit user override)"}
# --- gate 3: restart-required parameters are never applied live ----
if spec.change_class == ChangeClass.RESTART_REQUIRED or spec.applied_via == "restart":
p.status = "needs_restart"
rec = self._ledger(p, "needs_restart",
result=spec.restart_instruction or "restart required")
return {"ok": True, "status": "needs_restart",
"instruction": spec.restart_instruction,
"change_id": rec.id,
"note": "LPU never applies restart-required parameters to a "
"live process; follow the instruction and restart"}
# --- gate 4: level-1 auto-apply only when configured + SDK present ---
auto = (spec.level == 1 and self.auto_apply_level1
and not (user_confirmed or p.user_confirmed))
if self._sender is None:
p.status = "needs_sdk"
rec = self._ledger(p, "needs_sdk",
result="no command channel: apply manually in the "
"project (change is journaled here)")
return {"ok": True, "status": "needs_sdk", "change_id": rec.id,
"note": "no Integration SDK attached - change recorded as "
"a manual instruction, NOT applied"}
boundary = spec.change_class == ChangeClass.SAFE_BOUNDARY
extra = {"proposal_id": p.id, "level": int(spec.level),
"reason": p.reason}
result = self._sender(p.param, p.proposed, boundary, extra)
status = result.get("status", "failed")
if status not in APPLY_STATUSES:
status = "failed"
p.status = status
if status == "applied":
p.user_confirmed = True
rec = self._ledger(p, status, result=str(result.get("detail", "")))
return {"ok": status in ("applied", "queued_at_boundary", "needs_sdk",
"needs_restart"),
"status": status, "detail": result.get("detail", ""),
"change_id": rec.id, "auto_applied": auto}
# ----------------------------------------------------------------- rollback
def mark_boundary_applied(self, param: str, value) -> Optional[ChangeRecord]:
"""A safe-boundary change the SDK earlier queued has now crossed its
boundary (epoch/update edge): flip the ledger record to applied."""
for rec in reversed(self.ledger):
if (rec.param == param and rec.new_value == value
and rec.status == "queued_at_boundary"):
rec.status = "applied"
rec.applied_ts = time.time()
rec.result = (rec.result + " | boundary crossed, SDK acked applied").strip(" |")
self._write(rec)
prop = self.proposals.get(rec.proposal_id)
if prop:
prop.status = "applied"
return rec
return None
def rollback(self, change_id: str) -> dict:
rec = next((c for c in self.ledger if c.id == change_id), None)
if rec is None:
return {"ok": False, "error": "unknown change id"}
if rec.status == "rolled_back":
return {"ok": False, "error": "change already rolled back"}
if rec.status not in ("applied", "queued_at_boundary"):
return {"ok": False, "error": f"status {rec.status!r} cannot be "
"rolled back (nothing was applied)"}
if self._sender is None:
rec.status = "rolled_back"
rec.rollback_ts = time.time()
rec.rollback_result = "manual rollback required (no SDK channel)"
self._write(rec)
return {"ok": True, "status": "rolled_back",
"note": "no SDK channel: verify the manual rollback"}
spec = PARAM_SPECS[rec.param]
boundary = spec.change_class == ChangeClass.SAFE_BOUNDARY
result = self._sender(rec.param, rec.old_value, boundary,
{"rollback_of": rec.id, "reason": "rollback"})
ok = result.get("status") in ("applied", "queued_at_boundary")
rec.status = "rolled_back"
rec.rollback_ts = time.time()
rec.rollback_result = str(result.get("detail", result.get("status", "")))
self._write(rec)
prop = self.proposals.get(rec.proposal_id)
if prop:
prop.status = "rolled_back"
return {"ok": ok, "status": rec.status, "detail": rec.rollback_result}
# ------------------------------------------------------------------- ledger
def _ledger(self, p: Proposal, status: str, result: str = "") -> ChangeRecord:
rec = ChangeRecord(
id=self._new_id("chg"), proposal_id=p.id, param=p.param,
old_value=p.current, new_value=p.proposed, level=p.level,
reason=p.reason, status=status,
applied_ts=time.time() if status in ("applied", "queued_at_boundary",
"needs_restart", "needs_sdk") else None,
result=result,
)
self.ledger.append(rec)
self._write(rec)
return rec
def _write(self, rec: ChangeRecord) -> None:
if self._ledger_writer is not None:
try:
self._ledger_writer(rec.as_dict())
except Exception: # noqa: BLE001 - ledger write must not break control
pass
# ------------------------------------------------------------------ reports
def proposals_view(self) -> List[dict]:
return [p.as_dict() for p in
sorted(self.proposals.values(), key=lambda x: -x.ts)]
def ledger_view(self, limit: int = 50) -> List[dict]:
return [c.as_dict() for c in self.ledger[-limit:]]
|