File size: 27,292 Bytes
97eb21c | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 | """Evaluator bridge for declarative simulator-oracle patch programs.
The patch program is task-independent. A task adapter supplies only typed,
batched oracle signals; it does not choose a failure category, candidate, or
promotion outcome. This keeps environment plumbing separate from the open
category/program search.
"""
from __future__ import annotations
from collections.abc import Mapping
import copy
import math
import re
from typing import Any, Protocol, runtime_checkable
import torch
from .program import CompiledPatchProgram, PatchProgramError, compile_patch_program
class OracleProgramRuntimeError(RuntimeError):
"""Raised when an adapter or declarative program violates its contract."""
@runtime_checkable
class OracleSignalProvider(Protocol):
"""Minimal task-adapter surface required by the generic runtime."""
@property
def signal_specs(self) -> Mapping[str, Mapping[str, Any]]: ...
@property
def action_space(self) -> Mapping[str, Any]: ...
@property
def maximum_action_delta(self) -> tuple[float, ...]: ...
@property
def component_norms(self) -> tuple[Mapping[str, Any], ...]: ...
@property
def semantics_digest(self) -> str: ...
@property
def provenance(self) -> Mapping[str, Any]: ...
def bind_environment(self, env: Any) -> None: ...
def reset(self, batch_size: int) -> None: ...
def observe(
self,
*,
proposed_policy_action: torch.Tensor,
robot_qpos: torch.Tensor | None,
robot_qvel: torch.Tensor | None,
step: int,
) -> Mapping[str, torch.Tensor]: ...
class _OracleEstimatorSentinel:
"""Compatibility marker: this edition never instantiates an estimator."""
requires_policy_depth = False
direct_candidate_only = True
class SimOracleProgramPatcher:
"""Run one compiled finite program over a vectorized simulator batch."""
estimator = _OracleEstimatorSentinel()
def __init__(
self,
program: CompiledPatchProgram | Mapping[str, Any],
signal_provider: OracleSignalProvider,
) -> None:
if not isinstance(signal_provider, OracleSignalProvider):
raise OracleProgramRuntimeError(
"signal_provider does not implement OracleSignalProvider"
)
try:
# Always create a fresh immutable compilation bound to the exact
# live adapter. Never retain caller-owned compiled mappings.
self.program = compile_patch_program(
(
program.to_document()
if isinstance(program, CompiledPatchProgram)
else program
),
adapter_signal_specs=signal_provider.signal_specs,
)
except PatchProgramError as exc:
raise OracleProgramRuntimeError(str(exc)) from exc
expected_action_space = self.program.action_space.to_document()
if dict(signal_provider.action_space) != expected_action_space:
raise OracleProgramRuntimeError(
"program action space does not match the live task adapter"
)
maximum_action_delta = tuple(signal_provider.maximum_action_delta)
if (
len(maximum_action_delta) != self.program.action_space.dimension
or any(
isinstance(value, bool) or not isinstance(value, (int, float))
for value in maximum_action_delta
)
or any(
not math.isfinite(float(value)) or float(value) < 0.0
for value in maximum_action_delta
)
or any(
candidate > allowed
for candidate, allowed in zip(
self.program.limits.max_action_delta, maximum_action_delta
)
)
):
raise OracleProgramRuntimeError(
"program action-delta authority exceeds the live task adapter"
)
authored_from = self.program.to_document().get("authored_from")
if (
not isinstance(authored_from, Mapping)
or authored_from.get("semantics_digest")
!= signal_provider.semantics_digest
):
raise OracleProgramRuntimeError(
"program semantics identity does not match the live task adapter"
)
self._component_norms = tuple(
copy.deepcopy(dict(item)) for item in signal_provider.component_norms
)
for item in self._component_norms:
if set(item) != {"id", "indices", "maximum"}:
raise OracleProgramRuntimeError("live component norm schema is invalid")
indices = item["indices"]
maximum = item["maximum"]
if (
not isinstance(indices, (list, tuple))
or not indices
or tuple(sorted(indices)) != tuple(indices)
or len(set(indices)) != len(indices)
or any(
isinstance(index, bool)
or not isinstance(index, int)
or index < 0
or index >= self.program.action_space.dimension
for index in indices
)
or isinstance(maximum, bool)
or not isinstance(maximum, (int, float))
or not math.isfinite(float(maximum))
or maximum <= 0
or not isinstance(item["id"], str)
or re.fullmatch(r"[A-Za-z][A-Za-z0-9_.-]{0,127}", item["id"])
is None
):
raise OracleProgramRuntimeError("live component norm schema is invalid")
self.signal_provider = signal_provider
self._batch_size: int | None = None
self._runtimes: list[Any] = []
self._trace: list[dict[str, torch.Tensor]] = []
self._phase_indices = {
name: index for index, name in enumerate(self.program.phases)
}
@property
def provenance(self) -> dict[str, Any]:
return {
"schema_version": "sim_oracle_program_runtime.v1",
"runtime": "simulator_oracle",
"sim_only": True,
"deployment_eligible": False,
"metric_estimators_used": False,
"arbitrary_authored_code": False,
"program_id": self.program.id,
"program_sha256": self.program.sha256,
"program": self.program.to_document(),
"signal_provider": dict(self.signal_provider.provenance),
}
def bind_environment(self, env: Any) -> None:
self.signal_provider.bind_environment(env)
def reset(self, batch_size: int) -> None:
if isinstance(batch_size, bool) or not isinstance(batch_size, int) or batch_size < 1:
raise OracleProgramRuntimeError("batch_size must be a positive integer")
self._batch_size = None
self._runtimes = []
self._trace = []
self.signal_provider.reset(batch_size)
runtimes = [self.program.start_episode() for _ in range(batch_size)]
self._batch_size = batch_size
self._runtimes = runtimes
def _normalize_signals(
self,
signals: Mapping[str, torch.Tensor],
*,
batch_size: int,
) -> dict[str, torch.Tensor]:
if not isinstance(signals, Mapping):
raise OracleProgramRuntimeError("signal provider returned a non-mapping")
result: dict[str, torch.Tensor] = {}
for name, spec in self.program.signals.items():
value = signals.get(name)
if not isinstance(value, torch.Tensor):
raise OracleProgramRuntimeError(
f"signal provider did not return tensor {name!r}"
)
expected = (batch_size,) if spec.width == 1 else (batch_size, spec.width)
if tuple(value.shape) != expected:
raise OracleProgramRuntimeError(
f"signal {name!r} has shape {tuple(value.shape)}, expected {expected}"
)
if spec.value_type == "boolean":
if value.dtype != torch.bool:
raise OracleProgramRuntimeError(
f"signal {name!r} must have Boolean dtype"
)
elif not value.dtype.is_floating_point or not torch.isfinite(value).all():
raise OracleProgramRuntimeError(
f"signal {name!r} must have finite floating dtype"
)
result[name] = value
return result
def apply(
self,
*,
proposed_policy_action: torch.Tensor,
robot_qpos: torch.Tensor | None = None,
robot_qvel: torch.Tensor | None = None,
step: int,
**_: Any,
) -> torch.Tensor:
if self._batch_size is None or len(self._runtimes) != self._batch_size:
raise OracleProgramRuntimeError("program patcher must be reset before apply")
if (
proposed_policy_action.ndim != 2
or proposed_policy_action.shape[0] != self._batch_size
or proposed_policy_action.shape[1] != self.program.action_space.dimension
or not proposed_policy_action.dtype.is_floating_point
or not torch.isfinite(proposed_policy_action).all()
):
raise OracleProgramRuntimeError(
"proposed action does not match the compiled action space"
)
observed = self.signal_provider.observe(
proposed_policy_action=proposed_policy_action,
robot_qpos=robot_qpos,
robot_qvel=robot_qvel,
step=step,
)
signals = self._normalize_signals(observed, batch_size=self._batch_size)
policy_cpu = proposed_policy_action.detach().cpu().numpy()
signal_cpu = {
name: value.detach().cpu().numpy() for name, value in signals.items()
}
results = []
try:
for row_index, runtime in enumerate(self._runtimes):
row_signals = {
name: value[row_index] for name, value in signal_cpu.items()
}
results.append(
runtime.step(
policy_cpu[row_index],
row_signals,
step_index=step,
)
)
except PatchProgramError as exc:
raise OracleProgramRuntimeError(str(exc)) from exc
executed = torch.as_tensor(
[result.action.tolist() for result in results],
device=proposed_policy_action.device,
dtype=proposed_policy_action.dtype,
)
additive_selected = torch.as_tensor(
[result.trace["additive_selected"] for result in results],
device=executed.device,
dtype=torch.bool,
)
replacement_selected = torch.as_tensor(
[result.trace["replacement_selected"] for result in results],
device=executed.device,
dtype=torch.bool,
)
replacement_owned = torch.as_tensor(
[result.trace["replacement_owned"] for result in results],
device=executed.device,
dtype=torch.bool,
)
if torch.any(replacement_owned & ~replacement_selected):
raise OracleProgramRuntimeError(
"program trace owns a replacement without selecting one"
)
canonical_lower = torch.as_tensor(
self.program.action_space.lower,
device=executed.device,
dtype=torch.float64,
)
canonical_upper = torch.as_tensor(
self.program.action_space.upper,
device=executed.device,
dtype=torch.float64,
)
canonical_max_delta = torch.as_tensor(
self.program.limits.max_action_delta,
device=executed.device,
dtype=torch.float64,
)
replacement_anchor = (
(canonical_lower + canonical_upper) * 0.5
).to(dtype=executed.dtype)
requested_coordinates = additive_selected | replacement_selected
additive_owned = additive_selected & ~replacement_owned
def authority_values(current: torch.Tensor) -> torch.Tensor:
delta = current.to(dtype=torch.float64) - proposed_policy_action.to(
dtype=torch.float64
)
return torch.where(
replacement_owned,
current.to(dtype=torch.float64),
torch.where(additive_owned, delta, torch.zeros_like(delta)),
)
# The scalar interpreter computes in float64. Casting either an
# additive delta or an explicit replacement endpoint back to float32
# can round one ULP outside the canonical authority interval. Move
# only patch-owned coordinates inward. Additive edits move toward the
# incumbent action; replacements move toward the envelope midpoint.
for _ in range(8):
executed_canonical = executed.to(dtype=torch.float64)
proposed_canonical = proposed_policy_action.to(dtype=torch.float64)
authority = authority_values(executed)
value_excess = requested_coordinates & (
(authority < canonical_lower) | (authority > canonical_upper)
)
delta_excess = additive_owned & (
torch.abs(executed_canonical - proposed_canonical)
> canonical_max_delta
)
rounded_excess = value_excess | delta_excess
if not torch.any(rounded_excess):
break
target = torch.where(
replacement_owned & value_excess,
replacement_anchor.expand_as(executed),
proposed_policy_action,
)
executed = torch.where(
rounded_excess,
torch.nextafter(executed, target),
executed,
)
executed_canonical = executed.to(dtype=torch.float64)
proposed_canonical = proposed_policy_action.to(dtype=torch.float64)
authority = authority_values(executed)
rounded_excess = requested_coordinates & (
(authority < canonical_lower)
| (authority > canonical_upper)
| (
additive_owned
& (
torch.abs(executed_canonical - proposed_canonical)
> canonical_max_delta
)
)
)
executed = torch.where(
rounded_excess,
proposed_policy_action,
executed,
)
conversion_authority_clipped = torch.any(rounded_excess, dim=1)
# A failed conversion hands that coordinate back to the policy. A
# numerically unchanged but otherwise legal replacement retains
# ownership and remains part of its absolute component envelope.
replacement_owned &= ~rounded_excess
modified_coordinates = (
executed != proposed_policy_action
) & requested_coordinates
additive_owned &= modified_coordinates
authority_coordinates = replacement_owned | additive_owned
executed_canonical = executed.to(dtype=torch.float64)
proposed_canonical = proposed_policy_action.to(dtype=torch.float64)
authority = authority_values(executed)
if (
not torch.isfinite(executed).all()
or torch.any(authority_coordinates & (authority < canonical_lower))
or torch.any(authority_coordinates & (authority > canonical_upper))
or torch.any(
additive_owned
& (
torch.abs(executed_canonical - proposed_canonical)
> canonical_max_delta
)
)
):
raise OracleProgramRuntimeError(
"executed action is invalid after conversion to the live action dtype"
)
component_authority_clipped = torch.zeros(
executed.shape[0], device=executed.device, dtype=torch.bool
)
for item in self._component_norms:
indices = torch.as_tensor(item["indices"], device=executed.device)
component_active = torch.any(
authority_coordinates.index_select(1, indices), dim=1
)
# As with scalar endpoints, a vector exactly on a canonical norm
# boundary can round one ULP outward in the live dtype. Normalize
# representational excess inward along program-owned coordinates.
for _ in range(8):
authority = authority_values(executed)
norm_excess = component_active & (
torch.linalg.vector_norm(
authority.index_select(1, indices),
dim=1,
)
> float(item["maximum"])
)
if not torch.any(norm_excess):
break
current = executed.index_select(1, indices)
baseline = proposed_policy_action.index_select(1, indices)
owned = authority_coordinates.index_select(1, indices)
# Representation normalization may refine a real patch edit,
# but it must never turn a selected numerical no-op into a new
# intervention merely to shrink an absolute component vector.
modifiable = modified_coordinates.index_select(1, indices)
replacement = replacement_owned.index_select(1, indices)
inward_target = torch.where(
replacement,
replacement_anchor.index_select(0, indices).expand_as(current),
baseline,
)
inward = torch.nextafter(current, inward_target)
inward_canonical = inward.to(dtype=torch.float64)
baseline_canonical = baseline.to(dtype=torch.float64)
inward_authority = torch.where(
replacement,
inward_canonical,
inward_canonical - baseline_canonical,
)
scalar_conformant = (~owned) | (
(
inward_authority
>= canonical_lower.index_select(0, indices)
)
& (
inward_authority
<= canonical_upper.index_select(0, indices)
)
& (
replacement
| (
torch.abs(inward_canonical - baseline_canonical)
<= canonical_max_delta.index_select(0, indices)
)
)
)
current = torch.where(
norm_excess[:, None] & modifiable & scalar_conformant,
inward,
current,
)
executed[:, indices] = current
modified_coordinates = (
executed != proposed_policy_action
) & requested_coordinates
additive_owned = (
additive_selected & ~replacement_owned & modified_coordinates
)
authority_coordinates = replacement_owned | additive_owned
component_active = torch.any(
authority_coordinates.index_select(1, indices), dim=1
)
# Component envelopes may overlap. Abandoning a material edit in a
# later component can restore an out-of-envelope baseline coordinate
# and thereby invalidate a component checked earlier. Resolve that
# dependency to a fixed point. Every nonterminal pass removes at
# least one program-owned coordinate, so the action dimension is a
# strict deterministic convergence bound.
for _ in range(self.program.action_space.dimension):
modified_coordinates = (
executed != proposed_policy_action
) & requested_coordinates
additive_owned = (
additive_selected & ~replacement_owned & modified_coordinates
)
authority_coordinates = replacement_owned | additive_owned
authority = authority_values(executed)
abandon_coordinates = torch.zeros_like(authority_coordinates)
violating_rows = torch.zeros(
executed.shape[0], device=executed.device, dtype=torch.bool
)
for item in self._component_norms:
indices = torch.as_tensor(item["indices"], device=executed.device)
owned = authority_coordinates.index_select(1, indices)
norm_excess = torch.any(owned, dim=1) & (
torch.linalg.vector_norm(
authority.index_select(1, indices),
dim=1,
)
> float(item["maximum"])
)
if torch.any(norm_excess):
violating_rows |= norm_excess
abandon_coordinates[:, indices] |= norm_excess[:, None] & owned
if not torch.any(violating_rows):
break
before = int(authority_coordinates.sum().item())
executed = torch.where(
abandon_coordinates,
proposed_policy_action,
executed,
)
replacement_owned &= ~abandon_coordinates
component_authority_clipped |= violating_rows
modified_after = (
executed != proposed_policy_action
) & requested_coordinates
additive_after = (
additive_selected & ~replacement_owned & modified_after
)
after = int((replacement_owned | additive_after).sum().item())
if after >= before:
raise OracleProgramRuntimeError(
"component-norm fixed point made no conservative progress"
)
modified_coordinates = (
executed != proposed_policy_action
) & requested_coordinates
additive_owned = additive_selected & ~replacement_owned & modified_coordinates
authority_coordinates = replacement_owned | additive_owned
executed_canonical = executed.to(dtype=torch.float64)
proposed_canonical = proposed_policy_action.to(dtype=torch.float64)
authority = authority_values(executed)
if (
torch.any(authority_coordinates & (authority < canonical_lower))
or torch.any(authority_coordinates & (authority > canonical_upper))
or torch.any(
additive_owned
& (
torch.abs(executed_canonical - proposed_canonical)
> canonical_max_delta
)
)
):
raise OracleProgramRuntimeError(
"component normalization violated canonical scalar authority"
)
for item in self._component_norms:
indices = torch.as_tensor(item["indices"], device=executed.device)
remaining_excess = torch.any(
authority_coordinates.index_select(1, indices), dim=1
) & (
torch.linalg.vector_norm(
authority.index_select(1, indices),
dim=1,
)
> float(item["maximum"])
)
if torch.any(remaining_excess):
raise OracleProgramRuntimeError(
f"executed action violates live component norm {item['id']!r}"
)
modified = torch.any(executed != proposed_policy_action, dim=1)
scalar_modified = torch.as_tensor(
[bool(result.intervened) for result in results],
device=executed.device,
dtype=torch.bool,
)
if torch.any(modified & ~scalar_modified):
raise OracleProgramRuntimeError(
"live authority normalization created an undeclared intervention"
)
for runtime, was_modified, remains_modified in zip(
self._runtimes, scalar_modified.tolist(), modified.tolist()
):
if was_modified and not remains_modified:
if runtime.interventions < 1:
raise OracleProgramRuntimeError(
"runtime intervention accounting underflowed"
)
runtime.interventions -= 1
synchronized_counts = torch.as_tensor(
[runtime.interventions for runtime in self._runtimes],
device=executed.device,
dtype=torch.int64,
)
self._trace.append(
{
"modified": modified.detach().clone(),
"safety_clipped": torch.as_tensor(
[bool(result.trace["safety_clipped"]) for result in results],
device=executed.device,
dtype=torch.bool,
),
"authority_clipped": torch.as_tensor(
[bool(result.trace["authority_clipped"]) for result in results],
device=executed.device,
dtype=torch.bool,
) | component_authority_clipped | conversion_authority_clipped,
"additive_selected": additive_selected.detach().clone(),
"replacement_selected": replacement_selected.detach().clone(),
"replacement_owned": replacement_owned.detach().clone(),
"phase_before": torch.as_tensor(
[self._phase_indices[result.phase_before] for result in results],
device=executed.device,
dtype=torch.int64,
),
"phase_after": torch.as_tensor(
[self._phase_indices[result.phase_after] for result in results],
device=executed.device,
dtype=torch.int64,
),
"transitioned": torch.as_tensor(
[result.transition is not None for result in results],
device=executed.device,
dtype=torch.bool,
),
"intervention_count": synchronized_counts,
"proposed_action": proposed_policy_action.detach().clone(),
"executed_action": executed.detach().clone(),
}
)
return executed
def stacked_trace(self) -> dict[str, torch.Tensor]:
if not self._trace:
raise OracleProgramRuntimeError("no program trace has been recorded")
return {
name: torch.stack([row[name] for row in self._trace], dim=0)
for name in self._trace[0]
}
__all__ = [
"OracleProgramRuntimeError",
"OracleSignalProvider",
"SimOracleProgramPatcher",
]
|