@localmode/bench 0.3.1 → 0.5.0
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/README.md +48 -13
- package/dist/index.cjs +5 -5
- package/dist/index.cjs.map +1 -1
- package/dist/index.d.cts +158 -49
- package/dist/index.d.ts +158 -49
- package/dist/index.js +5 -5
- package/dist/index.js.map +1 -1
- package/package.json +1 -1
package/dist/index.d.ts
CHANGED
|
@@ -5,16 +5,20 @@
|
|
|
5
5
|
* published statistics can be recomputed and audited server-side.
|
|
6
6
|
*/
|
|
7
7
|
/** Protocol identifier embedded in every result. Bump only with a spec change. */
|
|
8
|
-
declare const BENCH_PROTOCOL_VERSION = "localmode-bench/
|
|
8
|
+
declare const BENCH_PROTOCOL_VERSION = "localmode-bench/3";
|
|
9
9
|
/** Result JSON schema version (independent of the protocol semantics version). */
|
|
10
10
|
declare const BENCH_SCHEMA_VERSION = 2;
|
|
11
11
|
/** Benchmark suite presets. `custom` = user-picked cells. */
|
|
12
12
|
type BenchSuiteId = 'quick' | 'standard' | 'thorough' | 'custom';
|
|
13
13
|
/**
|
|
14
14
|
* A benchmark runtime lane. Transformers.js is split into two lanes because
|
|
15
|
-
* WebGPU vs WASM is the
|
|
15
|
+
* WebGPU vs WASM is the core comparison axis.
|
|
16
16
|
*/
|
|
17
|
-
type BenchRuntimeId = 'transformers-webgpu' | 'transformers-wasm' | 'webllm'
|
|
17
|
+
type BenchRuntimeId = 'transformers-webgpu' | 'transformers-wasm' | 'webllm'
|
|
18
|
+
/** llama.cpp WASM on the CPU (`n_gpu_layers: 0`); before protocol v3 this lane silently ran on WebGPU where available. */
|
|
19
|
+
| 'wllama'
|
|
20
|
+
/** llama.cpp WASM with every layer offloaded to WebGPU. */
|
|
21
|
+
| 'wllama-webgpu' | 'litert' | 'chrome-ai' | 'mediapipe';
|
|
18
22
|
/** Workload kinds. Quality lanes are separate from timed performance lanes. */
|
|
19
23
|
type BenchWorkloadKind = 'llm-generate' | 'embed-single' | 'embed-batch' | 'quality-mmlu' | 'quality-sts';
|
|
20
24
|
/** A fixed LLM generation workload (public prompt, deterministic settings). */
|
|
@@ -169,25 +173,50 @@ interface BenchCellResult {
|
|
|
169
173
|
workloadKind: BenchWorkloadKind;
|
|
170
174
|
/** Backend actually used (probed, never the requested one). */
|
|
171
175
|
resolvedBackend: string;
|
|
176
|
+
/**
|
|
177
|
+
* Runtime configuration the adapter reports once the model is loaded
|
|
178
|
+
* (thread count, GPU layers requested and llama.cpp's offload report,
|
|
179
|
+
* dtype, ...): the equal-care record for the lane, per cell.
|
|
180
|
+
*/
|
|
181
|
+
runtimeConfig?: Record<string, string | number | boolean>;
|
|
172
182
|
load: LoadRecord | null;
|
|
173
183
|
/** Untimed warmup duration (ms), when a warmup ran. */
|
|
174
184
|
warmupMs?: number;
|
|
175
185
|
iterations: LLMIteration[] | EmbedIteration[];
|
|
186
|
+
/**
|
|
187
|
+
* Timed iterations the tab was hidden during, kept with their gates for
|
|
188
|
+
* auditability and repeated once the tab was visible again; never scored.
|
|
189
|
+
*/
|
|
190
|
+
discardedIterations?: Array<LLMIteration | EmbedIteration>;
|
|
176
191
|
memory?: MemorySample;
|
|
177
192
|
quality?: QualityResult;
|
|
178
193
|
status: BenchCellStatus;
|
|
179
194
|
invalidReasons?: string[];
|
|
195
|
+
/**
|
|
196
|
+
* Failed attempts that preceded the recorded outcome (a watchdog timeout, a
|
|
197
|
+
* provider error), oldest first. The runner retries a cell up to the
|
|
198
|
+
* policy's `maxAttempts`; nothing is retried silently.
|
|
199
|
+
*/
|
|
200
|
+
attempts?: Array<{
|
|
201
|
+
error: NonNullable<BenchCellResult['error']>;
|
|
202
|
+
at: number;
|
|
203
|
+
}>;
|
|
180
204
|
/** Error that ended the cell; `cause` carries the wrapped provider error's message when present. */
|
|
181
205
|
error?: {
|
|
182
206
|
name: string;
|
|
183
207
|
message: string;
|
|
208
|
+
/** Message of the wrapped provider error, when the thrown error carried a `cause`. */
|
|
184
209
|
cause?: string;
|
|
210
|
+
/** `name` of the wrapped provider error (e.g. wllama's `RuntimeError` for a WASM abort). */
|
|
211
|
+
causeName?: string;
|
|
212
|
+
/** Stack of the wrapped provider error, capped at 4,000 characters; a WASM abort names its frame only here. */
|
|
213
|
+
causeStack?: string;
|
|
185
214
|
};
|
|
186
215
|
}
|
|
187
216
|
/** Trace events global to the suite run (validity accounting). */
|
|
188
217
|
interface TraceEvent {
|
|
189
218
|
t: number;
|
|
190
|
-
type: 'suite-start' | 'suite-end' | 'visibility-hidden' | 'visibility-visible' | 'wakelock-acquired' | 'wakelock-released' | 'pressure-change' | 'gpu-device-lost' | 'cooldown-start' | 'cooldown-end' | 'abort';
|
|
219
|
+
type: 'suite-start' | 'suite-end' | 'visibility-hidden' | 'visibility-visible' | 'wakelock-acquired' | 'wakelock-released' | 'pressure-change' | 'gpu-device-lost' | 'cooldown-start' | 'cooldown-end' | 'cell-timeout' | 'cell-retry' | 'iteration-redo' | 'abort';
|
|
191
220
|
detail?: string;
|
|
192
221
|
}
|
|
193
222
|
/** Browser identification with its provenance. */
|
|
@@ -314,7 +343,7 @@ interface WasmFeatureSupport {
|
|
|
314
343
|
maxMemoryPages?: number;
|
|
315
344
|
}
|
|
316
345
|
/**
|
|
317
|
-
* Availability of the browser APIs the runtimes and the
|
|
346
|
+
* Availability of the browser APIs the runtimes and the analysis care about.
|
|
318
347
|
* Each entry is a plain presence check (the feature exists on this page),
|
|
319
348
|
* not a functional test. Chrome Built-in AI is reported by its
|
|
320
349
|
* `availability()` string where the API exists.
|
|
@@ -400,7 +429,7 @@ interface EnvironmentCapture {
|
|
|
400
429
|
browser: BrowserInfo;
|
|
401
430
|
os: OSInfo;
|
|
402
431
|
hardware: {
|
|
403
|
-
/** navigator.hardwareConcurrency —
|
|
432
|
+
/** navigator.hardwareConcurrency — WebKit clamps it (8 on macOS, 4 on iOS); Chromium and Firefox report the real count. */
|
|
404
433
|
cores: number | null;
|
|
405
434
|
coresClamped: boolean;
|
|
406
435
|
/** navigator.deviceMemory (GB) — Chromium-only, capped at 8. */
|
|
@@ -607,11 +636,17 @@ declare const WORKLOADS_BY_ID: ReadonlyMap<string, LLMWorkloadSpec | EmbedWorklo
|
|
|
607
636
|
/**
|
|
608
637
|
* Fixed runtime execution order (part of the versioned protocol). This makes
|
|
609
638
|
* the order deterministic so runtime interleaving is not a confounder across
|
|
610
|
-
* runs; it is a reproducibility measure, not a correctness fix.
|
|
611
|
-
*
|
|
612
|
-
*
|
|
613
|
-
*
|
|
614
|
-
*
|
|
639
|
+
* runs; it is a reproducibility measure, not a correctness fix. The order runs
|
|
640
|
+
* the WASM-arena runtimes first and the multi-GB-heap runtimes last; within a
|
|
641
|
+
* runtime, catalog order is preserved.
|
|
642
|
+
*
|
|
643
|
+
* The Transformers.js WASM lane runs before its WebGPU lane: Transformers.js
|
|
644
|
+
* serializes every ONNX session creation on one promise chain and never
|
|
645
|
+
* catches a rejection on it, so the first session that fails to create (a
|
|
646
|
+
* WebGPU execution provider the browser cannot initialize, an allocation
|
|
647
|
+
* failure under memory pressure) fails every later Transformers.js session in
|
|
648
|
+
* the page with the same error. Running the WASM lane first keeps a WebGPU
|
|
649
|
+
* failure from taking the CPU measurement with it.
|
|
615
650
|
*/
|
|
616
651
|
declare const RUNTIME_EXECUTION_ORDER: readonly BenchRuntimeId[];
|
|
617
652
|
/**
|
|
@@ -639,6 +674,39 @@ interface RunPolicy {
|
|
|
639
674
|
measureWarmReload: boolean;
|
|
640
675
|
/** CV above this fraction marks a summary metric as high-variance. */
|
|
641
676
|
highVarianceCv: number;
|
|
677
|
+
/**
|
|
678
|
+
* Watchdog: a load that reports no progress for this long is aborted as a
|
|
679
|
+
* timeout (stall-based, so a slow link keeps downloading as long as bytes
|
|
680
|
+
* arrive).
|
|
681
|
+
*/
|
|
682
|
+
loadStallMs: number;
|
|
683
|
+
/** Watchdog: absolute cap on one load attempt. */
|
|
684
|
+
loadTimeoutMs: number;
|
|
685
|
+
/**
|
|
686
|
+
* Watchdog: a generation (warmup or timed iteration) whose stream delivers
|
|
687
|
+
* nothing for this long is aborted as a timeout. A runtime surface that
|
|
688
|
+
* flushes every chunk in a terminal burst (LiteRT-LM) stays silent for the
|
|
689
|
+
* whole request, so this is well above a full request.
|
|
690
|
+
*/
|
|
691
|
+
chunkStallMs: number;
|
|
692
|
+
/** Watchdog: absolute cap on one generation or embedding call. */
|
|
693
|
+
iterationTimeoutMs: number;
|
|
694
|
+
/** Watchdog: absolute cap on one quality-lane cell (many generations). */
|
|
695
|
+
qualityTimeoutMs: number;
|
|
696
|
+
/**
|
|
697
|
+
* Attempts per cell (and per model load) before the cell is recorded as an
|
|
698
|
+
* error and the run moves on. Every failed attempt stays on the cell in
|
|
699
|
+
* `attempts`; retries are never silent.
|
|
700
|
+
*/
|
|
701
|
+
maxAttempts: number;
|
|
702
|
+
/**
|
|
703
|
+
* How long the runner waits for a hidden tab to become visible again before
|
|
704
|
+
* a timed iteration starts, or before it repeats an iteration the tab hid
|
|
705
|
+
* during. An iteration measured while hidden is kept on the cell in
|
|
706
|
+
* `discardedIterations` and repeated (up to `maxAttempts` times per
|
|
707
|
+
* iteration); a tab that stays hidden past this wait leaves the cell invalid.
|
|
708
|
+
*/
|
|
709
|
+
visibilityWaitMs: number;
|
|
642
710
|
}
|
|
643
711
|
/** Policies per suite preset. */
|
|
644
712
|
declare const RUN_POLICIES: Record<'quick' | 'standard' | 'thorough', RunPolicy>;
|
|
@@ -717,12 +785,15 @@ interface LoadedLLM {
|
|
|
717
785
|
model: BenchLanguageModel;
|
|
718
786
|
/** Backend actually in use — 'webgpu' | 'wasm' | 'gpu' | 'cpu' | 'chrome-builtin'. */
|
|
719
787
|
resolvedBackend: string;
|
|
788
|
+
/** Post-load runtime configuration worth recording on every cell (threads, GPU layers, dtype, ...). */
|
|
789
|
+
runtimeConfig?: Record<string, string | number | boolean>;
|
|
720
790
|
dispose(): Promise<void>;
|
|
721
791
|
}
|
|
722
792
|
/** A loaded embedding-model handle. */
|
|
723
793
|
interface LoadedEmbedder {
|
|
724
794
|
model: BenchEmbeddingModel;
|
|
725
795
|
resolvedBackend: string;
|
|
796
|
+
runtimeConfig?: Record<string, string | number | boolean>;
|
|
726
797
|
dispose(): Promise<void>;
|
|
727
798
|
}
|
|
728
799
|
/** Adapter for an LLM runtime lane. */
|
|
@@ -756,6 +827,37 @@ interface EmbeddingRuntimeAdapter {
|
|
|
756
827
|
/** Usage fidelity per runtime — how trustworthy provider token counts are. */
|
|
757
828
|
declare const USAGE_FIDELITY: Record<BenchRuntimeId, ProviderUsage['fidelity']>;
|
|
758
829
|
|
|
830
|
+
/**
|
|
831
|
+
* Suite-level trace recorder: validity-relevant events (tab visibility, wake
|
|
832
|
+
* lock, compute pressure, GPU device loss, aborts) with wall-clock timestamps.
|
|
833
|
+
* Iterations that overlap a hidden period are invalidated, never retried
|
|
834
|
+
* silently — the trace is part of the submitted, auditable record.
|
|
835
|
+
*/
|
|
836
|
+
|
|
837
|
+
/** Records trace events and tracks the current validity-gate state. */
|
|
838
|
+
declare class TraceRecorder {
|
|
839
|
+
private readonly events;
|
|
840
|
+
private disposers;
|
|
841
|
+
private hidden;
|
|
842
|
+
private lastPressureState;
|
|
843
|
+
private wakeLockSentinel;
|
|
844
|
+
/** Append an event at the current timestamp. */
|
|
845
|
+
record(type: TraceEvent['type'], detail?: string): void;
|
|
846
|
+
/** True while the page is hidden (timed regions overlapping this are invalid). */
|
|
847
|
+
get isHidden(): boolean;
|
|
848
|
+
/** Last observed compute-pressure state ('nominal'|'fair'|'serious'|'critical'). */
|
|
849
|
+
get pressureState(): string | undefined;
|
|
850
|
+
/** All recorded events (live reference; snapshot with `[...events]`). */
|
|
851
|
+
get all(): readonly TraceEvent[];
|
|
852
|
+
/**
|
|
853
|
+
* Attach browser listeners (visibility, compute pressure) and request a
|
|
854
|
+
* screen wake lock. Safe to call outside a browser (records nothing).
|
|
855
|
+
*/
|
|
856
|
+
attach(): Promise<void>;
|
|
857
|
+
/** Release listeners and the wake lock. */
|
|
858
|
+
dispose(): Promise<void>;
|
|
859
|
+
}
|
|
860
|
+
|
|
759
861
|
/**
|
|
760
862
|
* The suite runner: executes (runtime x model x workload) cells under the
|
|
761
863
|
* versioned run policy — cache probe, load, untimed warmup, N timed runs with
|
|
@@ -767,13 +869,43 @@ declare const USAGE_FIDELITY: Record<BenchRuntimeId, ProviderUsage['fidelity']>;
|
|
|
767
869
|
interface PlannedCell {
|
|
768
870
|
model: BenchModelRef;
|
|
769
871
|
workload: LLMWorkloadSpec | EmbedWorkloadSpec | QualityWorkloadSpec;
|
|
872
|
+
/**
|
|
873
|
+
* When set, the cell is recorded as `skipped` with this reason and never
|
|
874
|
+
* touches the adapter (a lane the submitter switched off, or a build the
|
|
875
|
+
* device cannot run). Keeps every cell a suite defines in the result.
|
|
876
|
+
*/
|
|
877
|
+
skipReason?: string;
|
|
878
|
+
}
|
|
879
|
+
/** A live-activity report: something observable happened inside a cell. */
|
|
880
|
+
interface RunnerActivity {
|
|
881
|
+
cellId: string;
|
|
882
|
+
/** `waiting-visible`: the tab is hidden and the runner is waiting for it to come back before timing. */
|
|
883
|
+
phase: 'load' | 'warmup' | 'iteration' | 'quality' | 'reload' | 'waiting-visible';
|
|
884
|
+
/** Current iteration or quality item (1-based) and the total, where applicable. */
|
|
885
|
+
iteration?: number;
|
|
886
|
+
total?: number;
|
|
887
|
+
/** Characters and chunks streamed so far in the current generation. */
|
|
888
|
+
chars?: number;
|
|
889
|
+
chunks?: number;
|
|
890
|
+
/** Load progress percentage, where the provider reports one. */
|
|
891
|
+
pct?: number;
|
|
770
892
|
}
|
|
771
893
|
/** Progress callbacks for a host UI. */
|
|
772
894
|
interface RunnerHooks {
|
|
895
|
+
/** The environment capture, before the fingerprint and the first cell (lets a host persist partial progress). */
|
|
896
|
+
onEnvironment?(environment: EnvironmentCapture): void;
|
|
773
897
|
onCellStart?(cellId: string, index: number, total: number): void;
|
|
774
898
|
onCellFinish?(cell: BenchCellResult): void;
|
|
775
899
|
onLoadProgress?(cellId: string, pct: number | undefined): void;
|
|
776
900
|
onIteration?(cellId: string, iteration: number, total: number): void;
|
|
901
|
+
/**
|
|
902
|
+
* Fires on every observable step inside a cell (each streamed chunk, each
|
|
903
|
+
* load progress event, each quality item), so a host can show that the run
|
|
904
|
+
* is alive and detect a stall at a glance.
|
|
905
|
+
*/
|
|
906
|
+
onActivity?(activity: RunnerActivity): void;
|
|
907
|
+
/** A cell attempt failed and the runner is about to retry it (`attempt` is the one starting, 2-based). */
|
|
908
|
+
onCellRetry?(cellId: string, attempt: number, error: NonNullable<BenchCellResult['error']>): void;
|
|
777
909
|
onPhase?(phase: string): void;
|
|
778
910
|
}
|
|
779
911
|
/** Inputs to a suite run. */
|
|
@@ -789,7 +921,11 @@ interface RunSuiteOptions {
|
|
|
789
921
|
/** Skip the fingerprint microbenchmark (tests only; submissions require it). */
|
|
790
922
|
skipFingerprint?: boolean;
|
|
791
923
|
userReportedDevice?: string;
|
|
924
|
+
/** Trace recorder to use instead of a fresh one (tests drive its visibility). */
|
|
925
|
+
trace?: TraceLike;
|
|
792
926
|
}
|
|
927
|
+
/** The trace recorder surface the runner needs. */
|
|
928
|
+
type TraceLike = Pick<TraceRecorder, 'attach' | 'dispose' | 'record' | 'isHidden' | 'pressureState' | 'all'>;
|
|
793
929
|
/**
|
|
794
930
|
* Run a benchmark suite and return the full, submittable result (raw traces
|
|
795
931
|
* included; `clientSummaries` computed with the same code the server uses).
|
|
@@ -857,12 +993,16 @@ interface DeviceTypeSignals {
|
|
|
857
993
|
formFactors?: string[];
|
|
858
994
|
/** UA-CH `mobile` bit. */
|
|
859
995
|
mobile?: boolean;
|
|
996
|
+
/** UA-CH `platform` (e.g. "Android"), which survives a "desktop site" UA rewrite. */
|
|
997
|
+
platform?: string;
|
|
860
998
|
}
|
|
861
999
|
/**
|
|
862
|
-
* Derive a form factor.
|
|
863
|
-
*
|
|
864
|
-
*
|
|
865
|
-
*
|
|
1000
|
+
* Derive a form factor. A mobile operating system settles it first: Android
|
|
1001
|
+
* and iOS devices are phones or tablets whatever the form-factor hint says (an
|
|
1002
|
+
* unfolded Galaxy Z Fold sends `formFactors: ["Desktop"]` with a tablet-style
|
|
1003
|
+
* UA). Then UA-CH form factors; then the UA, with `maxTouchPoints` unmasking
|
|
1004
|
+
* an iPad that reports itself as a Mac (iPadOS 13+ default) and separating
|
|
1005
|
+
* Android tablets (no `Mobile` token) from phones.
|
|
866
1006
|
*
|
|
867
1007
|
* @example
|
|
868
1008
|
* deriveDeviceType({ ua: navigator.userAgent, maxTouchPoints: navigator.maxTouchPoints });
|
|
@@ -960,37 +1100,6 @@ declare function runFingerprint(options?: {
|
|
|
960
1100
|
minDurationMs?: number;
|
|
961
1101
|
}): Promise<FingerprintResult>;
|
|
962
1102
|
|
|
963
|
-
/**
|
|
964
|
-
* Suite-level trace recorder: validity-relevant events (tab visibility, wake
|
|
965
|
-
* lock, compute pressure, GPU device loss, aborts) with wall-clock timestamps.
|
|
966
|
-
* Iterations that overlap a hidden period are invalidated, never retried
|
|
967
|
-
* silently — the trace is part of the submitted, auditable record.
|
|
968
|
-
*/
|
|
969
|
-
|
|
970
|
-
/** Records trace events and tracks the current validity-gate state. */
|
|
971
|
-
declare class TraceRecorder {
|
|
972
|
-
private readonly events;
|
|
973
|
-
private disposers;
|
|
974
|
-
private hidden;
|
|
975
|
-
private lastPressureState;
|
|
976
|
-
private wakeLockSentinel;
|
|
977
|
-
/** Append an event at the current timestamp. */
|
|
978
|
-
record(type: TraceEvent['type'], detail?: string): void;
|
|
979
|
-
/** True while the page is hidden (timed regions overlapping this are invalid). */
|
|
980
|
-
get isHidden(): boolean;
|
|
981
|
-
/** Last observed compute-pressure state ('nominal'|'fair'|'serious'|'critical'). */
|
|
982
|
-
get pressureState(): string | undefined;
|
|
983
|
-
/** All recorded events (live reference; snapshot with `[...events]`). */
|
|
984
|
-
get all(): readonly TraceEvent[];
|
|
985
|
-
/**
|
|
986
|
-
* Attach browser listeners (visibility, compute pressure) and request a
|
|
987
|
-
* screen wake lock. Safe to call outside a browser (records nothing).
|
|
988
|
-
*/
|
|
989
|
-
attach(): Promise<void>;
|
|
990
|
-
/** Release listeners and the wake lock. */
|
|
991
|
-
dispose(): Promise<void>;
|
|
992
|
-
}
|
|
993
|
-
|
|
994
1103
|
/**
|
|
995
1104
|
* Statistics for benchmark reporting: median headline, mean ± SD, IQR, 95% CI
|
|
996
1105
|
* (Student-t for the small n this protocol uses), CV, and geometric mean for
|
|
@@ -1130,7 +1239,7 @@ declare function computeRunDigest(result: BenchRunResult): Promise<string>;
|
|
|
1130
1239
|
declare function verifyRunDigest(result: BenchRunResult): Promise<boolean>;
|
|
1131
1240
|
|
|
1132
1241
|
/**
|
|
1133
|
-
* Aggregation for the public leaderboard and for
|
|
1242
|
+
* Aggregation for the public leaderboard and for offline analysis. Runs group
|
|
1134
1243
|
* into device-class rows; medians are taken per submission first, then across
|
|
1135
1244
|
* submissions (median-of-medians). Nothing is ever averaged across devices.
|
|
1136
1245
|
*/
|
|
@@ -1179,7 +1288,7 @@ declare function aggregateRuns(runs: readonly BenchRunResult[], minSubmissions?:
|
|
|
1179
1288
|
/** Leaderboard rows as CSV. */
|
|
1180
1289
|
declare function rowsToCSV(rows: readonly LeaderboardRow[]): string;
|
|
1181
1290
|
/**
|
|
1182
|
-
* Long-format per-iteration CSV for
|
|
1291
|
+
* Long-format per-iteration CSV for offline analysis (one row per timed
|
|
1183
1292
|
* iteration, with full environment identity columns) — feed to R/pandas.
|
|
1184
1293
|
*/
|
|
1185
1294
|
declare function runsToLongCSV(runs: readonly BenchRunResult[]): string;
|
|
@@ -1285,4 +1394,4 @@ interface STSPair {
|
|
|
1285
1394
|
/** 100-pair STS-B test subset (order preserved from the source dataset). */
|
|
1286
1395
|
declare const STSB_SUBSET: readonly STSPair[];
|
|
1287
1396
|
|
|
1288
|
-
export { type APIAvailability, type AdapterAvailability, type AdapterLoadProgress, BENCH_PROTOCOL_VERSION, BENCH_SCHEMA_VERSION, type BenchCellResult, type BenchCellStatus, type BenchChunk, type BenchEmbeddingModel, type BenchLanguageModel, type BenchModelRef, type BenchRunResult, type BenchRuntimeId, type BenchStreamChunk, type BenchSuiteId, type BenchWorkloadKind, type BenchWorkloadSpec, type BrowserInfo, type CellSummary, type DeviceInfo, type DeviceType, type DeviceTypeSignals, type DisplayInfo, EMBED_WORKLOADS, type EmbedIteration, type EmbedWorkloadSpec, type EmbeddingRuntimeAdapter, type EnvironmentCapture, type FingerprintResult, GENERATION_BUDGET, type GPUInfo, HEADLINE_MIN_SUBMISSIONS, type HarnessInfo, type LLMIteration, type LLMRuntimeAdapter, type LLMWorkloadSpec, LLM_WORKLOADS, type LeaderboardRow, type LoadRecord, type LoadedEmbedder, type LoadedLLM, type LocaleInfo, MIN_GENERATED_CHARS, type MMLUItem, MMLU_MAX_TOKENS, MMLU_OUTPUT_CAP, type MemorySample, type MetricSummary, type NetworkInfo, type OSInfo, PLAUSIBILITY_RULES_VERSION, type PlannedCell, type PlausibilityFlag, type ProviderUsage, QUALITY_WORKLOADS, type QualityResult, type QualityWorkloadSpec, RUNTIME_EXECUTION_ORDER, RUN_POLICIES, type RunPolicy, type RunSuiteOptions, type RunnerHooks, STREAM_COHERENCE_MIN_SPAN_RATIO, STSB_SUBSET, type STSPair, TINY_MMLU, type TraceEvent, TraceRecorder, USAGE_FIDELITY, type ValidationReport, WORKLOADS_BY_ID, type WasmFeatureSupport, type WebGLInfo, aggregateRuns, canonicalJson, captureEnvironment, checkPlausibility, computeRunDigest, deriveDeviceType, detectEngine, detectWasmFeatures, deviceClassOf, formatMMLUPrompt, geomean, hrNow, inferTimerResolutionUs, isIncrementalStream, mean, median, memoryApiAvailable, orderCells, parseGpuModel, parseMMLUAnswer, parseUserAgent, quantile, resolveGpuModel, rowsToCSV, runBenchmarkSuite, runFingerprint, runMMLUFidelity, runSTSQuality, runsToLongCSV, sampleMemoryBytes, sha256Hex, sleep, spearman, stddev, summarize, summarizeCell, summarizeRun, validateRunShape, validateSubmission, verifyRunDigest };
|
|
1397
|
+
export { type APIAvailability, type AdapterAvailability, type AdapterLoadProgress, BENCH_PROTOCOL_VERSION, BENCH_SCHEMA_VERSION, type BenchCellResult, type BenchCellStatus, type BenchChunk, type BenchEmbeddingModel, type BenchLanguageModel, type BenchModelRef, type BenchRunResult, type BenchRuntimeId, type BenchStreamChunk, type BenchSuiteId, type BenchWorkloadKind, type BenchWorkloadSpec, type BrowserInfo, type CellSummary, type DeviceInfo, type DeviceType, type DeviceTypeSignals, type DisplayInfo, EMBED_WORKLOADS, type EmbedIteration, type EmbedWorkloadSpec, type EmbeddingRuntimeAdapter, type EnvironmentCapture, type FingerprintResult, GENERATION_BUDGET, type GPUInfo, HEADLINE_MIN_SUBMISSIONS, type HarnessInfo, type LLMIteration, type LLMRuntimeAdapter, type LLMWorkloadSpec, LLM_WORKLOADS, type LeaderboardRow, type LoadRecord, type LoadedEmbedder, type LoadedLLM, type LocaleInfo, MIN_GENERATED_CHARS, type MMLUItem, MMLU_MAX_TOKENS, MMLU_OUTPUT_CAP, type MemorySample, type MetricSummary, type NetworkInfo, type OSInfo, PLAUSIBILITY_RULES_VERSION, type PlannedCell, type PlausibilityFlag, type ProviderUsage, QUALITY_WORKLOADS, type QualityResult, type QualityWorkloadSpec, RUNTIME_EXECUTION_ORDER, RUN_POLICIES, type RunPolicy, type RunSuiteOptions, type RunnerActivity, type RunnerHooks, STREAM_COHERENCE_MIN_SPAN_RATIO, STSB_SUBSET, type STSPair, TINY_MMLU, type TraceEvent, TraceRecorder, USAGE_FIDELITY, type ValidationReport, WORKLOADS_BY_ID, type WasmFeatureSupport, type WebGLInfo, aggregateRuns, canonicalJson, captureEnvironment, checkPlausibility, computeRunDigest, deriveDeviceType, detectEngine, detectWasmFeatures, deviceClassOf, formatMMLUPrompt, geomean, hrNow, inferTimerResolutionUs, isIncrementalStream, mean, median, memoryApiAvailable, orderCells, parseGpuModel, parseMMLUAnswer, parseUserAgent, quantile, resolveGpuModel, rowsToCSV, runBenchmarkSuite, runFingerprint, runMMLUFidelity, runSTSQuality, runsToLongCSV, sampleMemoryBytes, sha256Hex, sleep, spearman, stddev, summarize, summarizeCell, summarizeRun, validateRunShape, validateSubmission, verifyRunDigest };
|