@plasius/gpu-model-runtime 0.1.1 → 0.2.0

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/dist/index.cjs CHANGED
@@ -22,9 +22,12 @@ var index_exports = {};
22
22
  __export(index_exports, {
23
23
  AdapterRegistry: () => AdapterRegistry,
24
24
  MemoryModelCache: () => MemoryModelCache,
25
+ ModelResidencyError: () => ModelResidencyError,
26
+ ModelResidencyManager: () => ModelResidencyManager,
25
27
  ModelRuntime: () => ModelRuntime,
26
28
  PackageResourceResolver: () => PackageResourceResolver,
27
29
  createModelCacheKey: () => createModelCacheKey,
30
+ createModelResidencyKey: () => createModelResidencyKey,
28
31
  defaultSleep: () => defaultSleep,
29
32
  fetchResource: () => fetchResource,
30
33
  packageBootstrap: () => packageBootstrap,
@@ -543,6 +546,636 @@ function ensureHttpUrl(value) {
543
546
  return url.href;
544
547
  }
545
548
 
549
+ // src/model-residency.ts
550
+ var SHA256_PATTERN = /^[a-f0-9]{64}$/u;
551
+ var PARTITION_ID_PATTERN = /^[a-zA-Z0-9][a-zA-Z0-9._-]{0,127}$/u;
552
+ var ModelResidencyError = class extends Error {
553
+ code;
554
+ constructor(code, message) {
555
+ super(message);
556
+ this.name = "ModelResidencyError";
557
+ this.code = code;
558
+ }
559
+ };
560
+ function requireSafeNonNegativeInteger(value, name) {
561
+ if (!Number.isSafeInteger(value) || value < 0) {
562
+ throw new RangeError(`${name} must be a non-negative safe integer`);
563
+ }
564
+ }
565
+ function requirePositiveSafeInteger(value, name) {
566
+ if (!Number.isSafeInteger(value) || value <= 0) {
567
+ throw new RangeError(`${name} must be a positive safe integer`);
568
+ }
569
+ }
570
+ function requirePriority(priority) {
571
+ if (!Number.isFinite(priority) || priority < 0) {
572
+ throw new RangeError("priority must be a non-negative finite number");
573
+ }
574
+ }
575
+ function requireLod(lod) {
576
+ if (!Number.isInteger(lod) || lod < 0 || lod > 3) {
577
+ throw new RangeError("lod must be an integer from 0 to 3");
578
+ }
579
+ }
580
+ function createModelResidencyKey(assetRef, lod, partitionId) {
581
+ if (!SHA256_PATTERN.test(assetRef.contentHash)) {
582
+ throw new Error("ModelAssetRef.contentHash must be a lowercase SHA-256 value");
583
+ }
584
+ requireLod(lod);
585
+ if (partitionId !== void 0 && !PARTITION_ID_PATTERN.test(partitionId)) {
586
+ throw new Error("partitionId must be a bounded runtime token");
587
+ }
588
+ return `${assetRef.contentHash}:lod${lod}:${partitionId ?? "whole"}`;
589
+ }
590
+ function closedError() {
591
+ return new ModelResidencyError(
592
+ "manager-closed",
593
+ "The model residency manager is closed"
594
+ );
595
+ }
596
+ function abortedError() {
597
+ return new ModelResidencyError(
598
+ "acquire-aborted",
599
+ "The model residency acquisition was aborted"
600
+ );
601
+ }
602
+ function signalIsAborted(signal) {
603
+ return signal?.aborted === true;
604
+ }
605
+ var ModelResidencyManager = class {
606
+ #budget;
607
+ #load;
608
+ #loadTimeoutMs;
609
+ #entries = /* @__PURE__ */ new Map();
610
+ #loadTasks = /* @__PURE__ */ new Set();
611
+ #disposalTasks = /* @__PURE__ */ new Set();
612
+ #metrics = {
613
+ acquireCount: 0,
614
+ cacheHits: 0,
615
+ deduplicatedAcquires: 0,
616
+ loadsStarted: 0,
617
+ loadsCompleted: 0,
618
+ loadsFailed: 0,
619
+ loadsAborted: 0,
620
+ loadsTimedOut: 0,
621
+ evictions: 0,
622
+ disposals: 0,
623
+ disposalFailures: 0
624
+ };
625
+ #nextIdentity = 1;
626
+ #clock = 1;
627
+ #inFlightLoads = 0;
628
+ #closed = false;
629
+ #admissionTail = Promise.resolve();
630
+ #shutdownPromise;
631
+ constructor(options) {
632
+ requirePositiveSafeInteger(
633
+ options.budget.maxCpuBytes,
634
+ "budget.maxCpuBytes"
635
+ );
636
+ requirePositiveSafeInteger(
637
+ options.budget.maxGpuBytes,
638
+ "budget.maxGpuBytes"
639
+ );
640
+ requirePositiveSafeInteger(
641
+ options.budget.maxInFlightLoads,
642
+ "budget.maxInFlightLoads"
643
+ );
644
+ if (options.loadTimeoutMs !== void 0) {
645
+ requirePositiveSafeInteger(options.loadTimeoutMs, "loadTimeoutMs");
646
+ }
647
+ this.#budget = Object.freeze({ ...options.budget });
648
+ this.#load = options.load;
649
+ this.#loadTimeoutMs = options.loadTimeoutMs;
650
+ }
651
+ /** Acquire one shared resource lease, loading it at most once per cache key. */
652
+ acquire(options) {
653
+ try {
654
+ if (this.#closed) throw closedError();
655
+ requirePriority(options.priority);
656
+ requireSafeNonNegativeInteger(
657
+ options.estimatedCpuBytes,
658
+ "estimatedCpuBytes"
659
+ );
660
+ requireSafeNonNegativeInteger(
661
+ options.estimatedGpuBytes,
662
+ "estimatedGpuBytes"
663
+ );
664
+ const key = createModelResidencyKey(
665
+ options.assetRef,
666
+ options.lod,
667
+ options.partitionId
668
+ );
669
+ if (signalIsAborted(options.signal)) throw abortedError();
670
+ this.#metrics.acquireCount += 1;
671
+ const existing = this.#entries.get(key);
672
+ if (existing !== void 0) return this.#acquireExisting(existing, options);
673
+ return this.#scheduleAdmission(key, options);
674
+ } catch (error) {
675
+ return Promise.reject(error);
676
+ }
677
+ }
678
+ #acquireExisting(entry, options) {
679
+ this.#metrics.deduplicatedAcquires += 1;
680
+ if (entry.state === "resident") {
681
+ this.#metrics.cacheHits += 1;
682
+ return Promise.resolve(
683
+ this.#createLease(
684
+ entry,
685
+ options.priority,
686
+ options.pinned ?? false
687
+ )
688
+ );
689
+ }
690
+ return this.#addWaiter(entry, options);
691
+ }
692
+ #scheduleAdmission(key, options) {
693
+ let resolveAcquisition;
694
+ let rejectAcquisition;
695
+ const acquisition = new Promise(
696
+ (resolve, reject) => {
697
+ resolveAcquisition = resolve;
698
+ rejectAcquisition = reject;
699
+ }
700
+ );
701
+ const admission = this.#admissionTail.then(async () => {
702
+ if (this.#closed) throw closedError();
703
+ if (signalIsAborted(options.signal)) throw abortedError();
704
+ const existing = this.#entries.get(key);
705
+ if (existing !== void 0) {
706
+ void this.#acquireExisting(existing, options).then(
707
+ resolveAcquisition,
708
+ rejectAcquisition
709
+ );
710
+ return;
711
+ }
712
+ await this.#ensureCapacityForReservation(
713
+ options.estimatedCpuBytes,
714
+ options.estimatedGpuBytes
715
+ );
716
+ if (this.#closed) throw closedError();
717
+ if (signalIsAborted(options.signal)) throw abortedError();
718
+ const entry = {
719
+ key,
720
+ assetRef: options.assetRef,
721
+ lod: options.lod,
722
+ ...options.partitionId === void 0 ? {} : { partitionId: options.partitionId },
723
+ estimatedCpuBytes: options.estimatedCpuBytes,
724
+ estimatedGpuBytes: options.estimatedGpuBytes,
725
+ queueOrder: this.#nextIdentity++,
726
+ state: "queued",
727
+ priority: options.priority,
728
+ retainedPriority: options.priority,
729
+ lastUsed: this.#clock++,
730
+ abandoned: false,
731
+ loadSlotReleased: true,
732
+ waiters: /* @__PURE__ */ new Map(),
733
+ leases: /* @__PURE__ */ new Map()
734
+ };
735
+ this.#entries.set(key, entry);
736
+ void this.#addWaiter(entry, options).then(
737
+ resolveAcquisition,
738
+ rejectAcquisition
739
+ );
740
+ this.#pumpQueue();
741
+ });
742
+ this.#admissionTail = admission.catch(() => void 0);
743
+ void admission.catch(rejectAcquisition);
744
+ return acquisition;
745
+ }
746
+ /** Dispose every currently unreferenced, unpinned resident cache entry. */
747
+ async evictUnused() {
748
+ const tasks = [];
749
+ for (const entry of this.#evictionCandidates()) {
750
+ tasks.push(this.#evict(entry));
751
+ }
752
+ await Promise.all(tasks);
753
+ }
754
+ /** Wait until queued/loading work and its resulting disposal have settled. */
755
+ async waitForIdle() {
756
+ await this.#admissionTail;
757
+ do {
758
+ this.#pumpQueue();
759
+ const tasks = [...this.#loadTasks];
760
+ if (tasks.length > 0) {
761
+ await Promise.allSettled(tasks);
762
+ }
763
+ } while (this.#loadTasks.size > 0 || this.#queuedCount() > 0);
764
+ await Promise.allSettled([...this.#disposalTasks]);
765
+ }
766
+ /** Abort pending work and dispose every resident or late-completing resource. */
767
+ shutdown() {
768
+ if (this.#shutdownPromise !== void 0) return this.#shutdownPromise;
769
+ this.#closed = true;
770
+ const error = closedError();
771
+ const disposalTasks = [];
772
+ for (const entry of [...this.#entries.values()]) {
773
+ if (entry.state === "resident") {
774
+ this.#entries.delete(entry.key);
775
+ disposalTasks.push(this.#disposeEntry(entry));
776
+ continue;
777
+ }
778
+ this.#rejectWaiters(entry, error);
779
+ if (entry.state === "queued") {
780
+ this.#entries.delete(entry.key);
781
+ } else {
782
+ this.#abandonLoadingEntry(entry, error, false);
783
+ }
784
+ }
785
+ this.#shutdownPromise = (async () => {
786
+ await Promise.allSettled(disposalTasks);
787
+ await this.waitForIdle();
788
+ await Promise.allSettled([...this.#disposalTasks]);
789
+ })();
790
+ return this.#shutdownPromise;
791
+ }
792
+ /** Return immutable byte, queue, cache, and lifetime metrics. */
793
+ snapshot() {
794
+ let cpuBytes = 0;
795
+ let gpuBytes = 0;
796
+ let reservedCpuBytes = 0;
797
+ let reservedGpuBytes = 0;
798
+ let referenceCount = 0;
799
+ let pinnedCount = 0;
800
+ let residentCount = 0;
801
+ let pendingCount = 0;
802
+ const entries = [];
803
+ for (const entry of this.#entries.values()) {
804
+ const pinned = [...entry.leases.values()].some(
805
+ (lease) => lease.active && lease.pinned
806
+ );
807
+ const references = [...entry.leases.values()].filter(
808
+ (lease) => lease.active
809
+ ).length;
810
+ referenceCount += references;
811
+ if (pinned) pinnedCount += 1;
812
+ if (entry.state === "resident") {
813
+ residentCount += 1;
814
+ cpuBytes += entry.loaded.cpuBytes;
815
+ gpuBytes += entry.loaded.gpuBytes;
816
+ } else {
817
+ pendingCount += 1;
818
+ reservedCpuBytes += entry.estimatedCpuBytes;
819
+ reservedGpuBytes += entry.estimatedGpuBytes;
820
+ }
821
+ entries.push(
822
+ Object.freeze({
823
+ key: entry.key,
824
+ state: entry.state,
825
+ priority: entry.priority,
826
+ references,
827
+ pinned,
828
+ cpuBytes: entry.state === "resident" ? entry.loaded.cpuBytes : entry.estimatedCpuBytes,
829
+ gpuBytes: entry.state === "resident" ? entry.loaded.gpuBytes : entry.estimatedGpuBytes
830
+ })
831
+ );
832
+ }
833
+ entries.sort((left, right) => left.key.localeCompare(right.key));
834
+ return Object.freeze({
835
+ ...this.#budget,
836
+ cpuBytes,
837
+ gpuBytes,
838
+ reservedCpuBytes,
839
+ reservedGpuBytes,
840
+ residentCount,
841
+ pendingCount,
842
+ referenceCount,
843
+ pinnedCount,
844
+ inFlightLoads: this.#inFlightLoads,
845
+ queuedLoads: this.#queuedCount(),
846
+ ...this.#metrics,
847
+ residentKeys: Object.freeze(
848
+ entries.filter((entry) => entry.state === "resident").map((entry) => entry.key)
849
+ ),
850
+ entries: Object.freeze(entries)
851
+ });
852
+ }
853
+ #addWaiter(entry, options) {
854
+ const id = this.#nextIdentity++;
855
+ let resolvePromise;
856
+ let rejectPromise;
857
+ const promise = new Promise(
858
+ (resolve, reject) => {
859
+ resolvePromise = resolve;
860
+ rejectPromise = reject;
861
+ }
862
+ );
863
+ const waiter = {
864
+ id,
865
+ priority: options.priority,
866
+ pinned: options.pinned ?? false,
867
+ ...options.signal === void 0 ? {} : { signal: options.signal },
868
+ resolve: resolvePromise,
869
+ reject: rejectPromise
870
+ };
871
+ if (options.signal !== void 0) {
872
+ waiter.onAbort = () => this.#abortWaiter(entry, waiter);
873
+ options.signal.addEventListener("abort", waiter.onAbort, { once: true });
874
+ }
875
+ entry.waiters.set(id, waiter);
876
+ entry.retainedPriority = Math.max(entry.retainedPriority, options.priority);
877
+ this.#recalculatePriority(entry);
878
+ return promise;
879
+ }
880
+ #abortWaiter(entry, waiter) {
881
+ entry.waiters.delete(waiter.id);
882
+ this.#detachWaiterSignal(waiter);
883
+ waiter.reject(abortedError());
884
+ this.#recalculatePriority(entry);
885
+ if (entry.waiters.size > 0) return;
886
+ if (entry.state === "queued") {
887
+ this.#entries.delete(entry.key);
888
+ this.#pumpQueue();
889
+ return;
890
+ }
891
+ this.#abandonLoadingEntry(entry, abortedError(), true);
892
+ }
893
+ #createLease(entry, priority, pinned) {
894
+ const id = this.#nextIdentity++;
895
+ const record = {
896
+ id,
897
+ active: true,
898
+ priority,
899
+ pinned
900
+ };
901
+ entry.leases.set(id, record);
902
+ entry.retainedPriority = priority;
903
+ entry.lastUsed = this.#clock++;
904
+ this.#recalculatePriority(entry);
905
+ const lease = {
906
+ key: entry.key,
907
+ assetRef: entry.assetRef,
908
+ lod: entry.lod,
909
+ ...entry.partitionId === void 0 ? {} : { partitionId: entry.partitionId },
910
+ resource: entry.loaded.resource,
911
+ release: () => {
912
+ if (!record.active) return;
913
+ record.active = false;
914
+ entry.leases.delete(id);
915
+ entry.lastUsed = this.#clock++;
916
+ this.#recalculatePriority(entry);
917
+ },
918
+ setPriority: (nextPriority) => {
919
+ if (!record.active || this.#entries.get(entry.key) !== entry) return;
920
+ requirePriority(nextPriority);
921
+ record.priority = nextPriority;
922
+ entry.retainedPriority = nextPriority;
923
+ entry.lastUsed = this.#clock++;
924
+ this.#recalculatePriority(entry);
925
+ },
926
+ setPinned: (nextPinned) => {
927
+ if (!record.active || this.#entries.get(entry.key) !== entry) return;
928
+ record.pinned = nextPinned;
929
+ }
930
+ };
931
+ return Object.freeze(lease);
932
+ }
933
+ #recalculatePriority(entry) {
934
+ let priority = entry.retainedPriority;
935
+ for (const waiter of entry.waiters.values()) {
936
+ priority = Math.max(priority, waiter.priority);
937
+ }
938
+ for (const lease of entry.leases.values()) {
939
+ priority = Math.max(priority, lease.priority);
940
+ }
941
+ entry.priority = priority;
942
+ }
943
+ #pumpQueue() {
944
+ if (this.#closed) return;
945
+ while (this.#inFlightLoads < this.#budget.maxInFlightLoads) {
946
+ const candidate = [...this.#entries.values()].filter(
947
+ (entry) => entry.state === "queued" && !entry.abandoned && entry.waiters.size > 0
948
+ ).sort(
949
+ (left, right) => right.priority - left.priority || left.queueOrder - right.queueOrder
950
+ )[0];
951
+ if (candidate === void 0) return;
952
+ this.#startLoad(candidate);
953
+ }
954
+ }
955
+ #startLoad(entry) {
956
+ entry.state = "loading";
957
+ entry.loadSlotReleased = false;
958
+ entry.controller = new AbortController();
959
+ this.#inFlightLoads += 1;
960
+ this.#metrics.loadsStarted += 1;
961
+ if (this.#loadTimeoutMs !== void 0) {
962
+ entry.timeout = setTimeout(() => {
963
+ this.#metrics.loadsTimedOut += 1;
964
+ this.#abandonLoadingEntry(
965
+ entry,
966
+ new ModelResidencyError(
967
+ "load-timeout",
968
+ `Model load exceeded ${this.#loadTimeoutMs} ms`
969
+ ),
970
+ true
971
+ );
972
+ }, this.#loadTimeoutMs);
973
+ }
974
+ let adapterPromise;
975
+ try {
976
+ adapterPromise = this.#load({
977
+ assetRef: entry.assetRef,
978
+ lod: entry.lod,
979
+ ...entry.partitionId === void 0 ? {} : { partitionId: entry.partitionId },
980
+ priority: entry.priority,
981
+ signal: entry.controller.signal
982
+ });
983
+ } catch (error) {
984
+ adapterPromise = Promise.reject(error);
985
+ }
986
+ const task = Promise.resolve(adapterPromise).then((result) => this.#completeLoad(entry, result)).catch((error) => this.#failLoad(entry, error)).finally(() => {
987
+ if (entry.timeout !== void 0) clearTimeout(entry.timeout);
988
+ entry.timeout = void 0;
989
+ entry.controller = void 0;
990
+ this.#releaseLoadSlot(entry);
991
+ entry.loadTask = void 0;
992
+ });
993
+ entry.loadTask = task;
994
+ this.#loadTasks.add(task);
995
+ }
996
+ async #completeLoad(entry, result) {
997
+ if (entry.abandoned) {
998
+ await this.#disposeLoaded(result);
999
+ return;
1000
+ }
1001
+ try {
1002
+ this.#validateLoadedResource(entry, result);
1003
+ const evictions = this.#ensureCapacityForLoadedResult(entry, result);
1004
+ await Promise.all(evictions);
1005
+ } catch (error) {
1006
+ await this.#disposeLoaded(result);
1007
+ this.#entries.delete(entry.key);
1008
+ this.#rejectWaiters(entry, error);
1009
+ this.#metrics.loadsFailed += 1;
1010
+ return;
1011
+ }
1012
+ entry.state = "resident";
1013
+ entry.loaded = result;
1014
+ entry.lastUsed = this.#clock++;
1015
+ this.#metrics.loadsCompleted += 1;
1016
+ const waiters = [...entry.waiters.values()];
1017
+ entry.waiters.clear();
1018
+ for (const waiter of waiters) {
1019
+ this.#detachWaiterSignal(waiter);
1020
+ waiter.resolve(this.#createLease(entry, waiter.priority, waiter.pinned));
1021
+ }
1022
+ }
1023
+ #failLoad(entry, error) {
1024
+ if (entry.abandoned) return;
1025
+ this.#entries.delete(entry.key);
1026
+ this.#metrics.loadsFailed += 1;
1027
+ this.#rejectWaiters(entry, error);
1028
+ }
1029
+ #abandonLoadingEntry(entry, error, rejectWaiters) {
1030
+ entry.abandoned = true;
1031
+ this.#metrics.loadsAborted += 1;
1032
+ if (entry.timeout !== void 0) clearTimeout(entry.timeout);
1033
+ entry.timeout = void 0;
1034
+ this.#entries.delete(entry.key);
1035
+ if (rejectWaiters) this.#rejectWaiters(entry, error);
1036
+ entry.controller?.abort(error);
1037
+ this.#releaseLoadSlot(entry);
1038
+ }
1039
+ #releaseLoadSlot(entry) {
1040
+ if (entry.loadSlotReleased) return;
1041
+ entry.loadSlotReleased = true;
1042
+ this.#inFlightLoads -= 1;
1043
+ this.#loadTasks.delete(entry.loadTask);
1044
+ this.#pumpQueue();
1045
+ }
1046
+ #rejectWaiters(entry, error) {
1047
+ const waiters = [...entry.waiters.values()];
1048
+ entry.waiters.clear();
1049
+ for (const waiter of waiters) {
1050
+ this.#detachWaiterSignal(waiter);
1051
+ waiter.reject(error);
1052
+ }
1053
+ this.#recalculatePriority(entry);
1054
+ }
1055
+ #detachWaiterSignal(waiter) {
1056
+ if (waiter.signal !== void 0 && waiter.onAbort !== void 0) {
1057
+ waiter.signal.removeEventListener("abort", waiter.onAbort);
1058
+ }
1059
+ waiter.onAbort = void 0;
1060
+ }
1061
+ #validateLoadedResource(entry, result) {
1062
+ if (result === null || typeof result !== "object") {
1063
+ throw new ModelResidencyError(
1064
+ "invalid-load-result",
1065
+ "The model loader returned an invalid resource"
1066
+ );
1067
+ }
1068
+ if (typeof result.dispose !== "function") {
1069
+ throw new ModelResidencyError(
1070
+ "invalid-load-result",
1071
+ "The model loader must return a disposer"
1072
+ );
1073
+ }
1074
+ requireSafeNonNegativeInteger(result.cpuBytes, "loaded.cpuBytes");
1075
+ requireSafeNonNegativeInteger(result.gpuBytes, "loaded.gpuBytes");
1076
+ if (result.contentHash !== entry.assetRef.contentHash) {
1077
+ throw new ModelResidencyError(
1078
+ "content-hash-mismatch",
1079
+ "The loaded model content hash does not match its canonical reference"
1080
+ );
1081
+ }
1082
+ }
1083
+ async #ensureCapacityForReservation(cpuBytes, gpuBytes) {
1084
+ if (cpuBytes > this.#budget.maxCpuBytes || gpuBytes > this.#budget.maxGpuBytes) {
1085
+ throw this.#budgetError();
1086
+ }
1087
+ let usage = this.#totalAccountedBytes();
1088
+ for (const candidate of this.#evictionCandidates()) {
1089
+ if (usage.cpu + cpuBytes <= this.#budget.maxCpuBytes && usage.gpu + gpuBytes <= this.#budget.maxGpuBytes) {
1090
+ return;
1091
+ }
1092
+ await this.#evict(candidate);
1093
+ usage = this.#totalAccountedBytes();
1094
+ }
1095
+ if (usage.cpu + cpuBytes > this.#budget.maxCpuBytes || usage.gpu + gpuBytes > this.#budget.maxGpuBytes) {
1096
+ throw this.#budgetError();
1097
+ }
1098
+ }
1099
+ #ensureCapacityForLoadedResult(loadingEntry, result) {
1100
+ const usage = this.#totalAccountedBytes();
1101
+ usage.cpu += result.cpuBytes - loadingEntry.estimatedCpuBytes;
1102
+ usage.gpu += result.gpuBytes - loadingEntry.estimatedGpuBytes;
1103
+ const disposals = [];
1104
+ for (const candidate of this.#evictionCandidates()) {
1105
+ if (usage.cpu <= this.#budget.maxCpuBytes && usage.gpu <= this.#budget.maxGpuBytes) {
1106
+ return disposals;
1107
+ }
1108
+ usage.cpu -= candidate.loaded.cpuBytes;
1109
+ usage.gpu -= candidate.loaded.gpuBytes;
1110
+ disposals.push(this.#evict(candidate));
1111
+ }
1112
+ if (usage.cpu > this.#budget.maxCpuBytes || usage.gpu > this.#budget.maxGpuBytes) {
1113
+ throw this.#budgetError();
1114
+ }
1115
+ return disposals;
1116
+ }
1117
+ #budgetError() {
1118
+ return new ModelResidencyError(
1119
+ "budget-exceeded",
1120
+ "Pinned, referenced, or pending models leave insufficient residency budget"
1121
+ );
1122
+ }
1123
+ #totalAccountedBytes() {
1124
+ let cpu = 0;
1125
+ let gpu = 0;
1126
+ for (const entry of this.#entries.values()) {
1127
+ if (entry.state === "resident") {
1128
+ cpu += entry.loaded.cpuBytes;
1129
+ gpu += entry.loaded.gpuBytes;
1130
+ } else {
1131
+ cpu += entry.estimatedCpuBytes;
1132
+ gpu += entry.estimatedGpuBytes;
1133
+ }
1134
+ }
1135
+ return { cpu, gpu };
1136
+ }
1137
+ #evictionCandidates() {
1138
+ return [...this.#entries.values()].filter(
1139
+ (entry) => entry.state === "resident" && entry.leases.size === 0
1140
+ ).sort(
1141
+ (left, right) => left.priority - right.priority || left.lastUsed - right.lastUsed
1142
+ );
1143
+ }
1144
+ #evict(entry) {
1145
+ this.#entries.delete(entry.key);
1146
+ this.#metrics.evictions += 1;
1147
+ return this.#disposeEntry(entry);
1148
+ }
1149
+ #disposeEntry(entry) {
1150
+ const result = entry.loaded;
1151
+ entry.loaded = void 0;
1152
+ return this.#disposeLoaded(result);
1153
+ }
1154
+ #disposeLoaded(result) {
1155
+ this.#metrics.disposals += 1;
1156
+ let disposal;
1157
+ try {
1158
+ disposal = Promise.resolve(result.dispose());
1159
+ } catch (error) {
1160
+ disposal = Promise.reject(error);
1161
+ }
1162
+ const tracked = disposal.catch(() => {
1163
+ this.#metrics.disposalFailures += 1;
1164
+ }).finally(() => {
1165
+ this.#disposalTasks.delete(tracked);
1166
+ });
1167
+ this.#disposalTasks.add(tracked);
1168
+ return tracked;
1169
+ }
1170
+ #queuedCount() {
1171
+ let count = 0;
1172
+ for (const entry of this.#entries.values()) {
1173
+ if (entry.state === "queued") count += 1;
1174
+ }
1175
+ return count;
1176
+ }
1177
+ };
1178
+
546
1179
  // src/index.ts
547
1180
  var packageName = "@plasius/gpu-model-runtime";
548
1181
  var packageBootstrap = Object.freeze({
@@ -554,9 +1187,12 @@ var packageBootstrap = Object.freeze({
554
1187
  0 && (module.exports = {
555
1188
  AdapterRegistry,
556
1189
  MemoryModelCache,
1190
+ ModelResidencyError,
1191
+ ModelResidencyManager,
557
1192
  ModelRuntime,
558
1193
  PackageResourceResolver,
559
1194
  createModelCacheKey,
1195
+ createModelResidencyKey,
560
1196
  defaultSleep,
561
1197
  fetchResource,
562
1198
  packageBootstrap,