lynae-1219 commited on
Commit
2e2d827
·
verified ·
1 Parent(s): d670737

Upload folder using huggingface_hub

Browse files
README.md CHANGED
@@ -167,8 +167,7 @@ dataset_info:
167
  dtype: float64
168
  splits:
169
  - name: summary
170
- num_examples: 223
171
- num_bytes: 70836
172
  - config_name: per_layer_kernel
173
  features:
174
  - name: record_type
@@ -309,7 +308,7 @@ dataset_info:
309
 
310
  # AgentPerfBench
311
 
312
- LLM inference benchmark: 3,155 main sweep rows and 37 per-layer kernel validation rows, plus 148,077 per-kernel NCU profiles, across 9 models, 14 GPU configurations, and 2 serving engines (vLLM 0.19.0, SGLang 0.5.9). All models served in BF16 except gpt-oss, which uses mxfp4 for projection weights.
313
 
314
  ## Dataset configurations
315
 
@@ -319,11 +318,11 @@ Replays exact ISL/OSL sequences from recorded agent sessions (SWE-Bench, Termina
319
 
320
  17 profiles: `chat-medium`, `chat-multiturn-long`, `chat-multiturn-medium`, `chat-multiturn-short`, `chat-short`, `chat-singleturn`, `coding-singleturn`, `decode-heavy`, `osworld-multiturn-long`, `osworld-multiturn-medium`, `osworld-multiturn-short`, `prefill-heavy`, `random-1k`, `swebench-multiturn-medium`, `swebench-multiturn-short`, `terminalbench-multiturn-medium`, `terminalbench-multiturn-short`
321
 
322
- ### synthetic_distributional (223 rows)
323
 
324
- ISL/OSL sampled from lognormal fits to real workload statistics. 46 unique (model, hardware, engine) combinations, 14 profiles, 3 concurrency levels {1, 5, 10}, 2.8% matrix fill.
325
 
326
- 14 profiles: `chat-medium`, `chat-multiturn`, `chat-multiturn-long`, `chat-multiturn-medium`, `chat-multiturn-short`, `chat-short`, `chat-singleturn`, `coding-singleturn`, `osworld-multiturn`, `osworld-multiturn-long`, `osworld-multiturn-medium`, `swebench-multiturn`, `terminalbench-multiturn`, `terminalbench-multiturn-short`
327
 
328
  ### per_layer_kernel (37 rows)
329
 
@@ -339,12 +338,12 @@ Curated H100 / Llama-3.1-8B / vLLM validation table for the distributional synth
339
 
340
  ### Quality filtering
341
 
342
- Concurrency levels: trace_replay {1, 5, 10, 20, 40, 80}, synthetic_distributional {1, 5, 10}. Configurations where fewer than 75% of requests completed successfully are excluded. Summary metrics are computed from successful requests only.
343
 
344
  | Config | Rows |
345
  |--------|------|
346
  | trace_replay | 2,932 |
347
- | synthetic_distributional | 223 |
348
  | per_layer_kernel | 37 |
349
  | kernels_labeled | 148,077 |
350
  | mse_validation | 28 |
 
167
  dtype: float64
168
  splits:
169
  - name: summary
170
+ num_examples: 1365
 
171
  - config_name: per_layer_kernel
172
  features:
173
  - name: record_type
 
308
 
309
  # AgentPerfBench
310
 
311
+ LLM inference benchmark: 4,297 main sweep rows and 37 per-layer kernel validation rows, plus 148,077 per-kernel NCU profiles, across 9 models, 14 GPU configurations, and 2 serving engines (vLLM 0.19.0, SGLang 0.5.9). All models served in BF16 except gpt-oss, which uses mxfp4 for projection weights.
312
 
313
  ## Dataset configurations
314
 
 
318
 
319
  17 profiles: `chat-medium`, `chat-multiturn-long`, `chat-multiturn-medium`, `chat-multiturn-short`, `chat-short`, `chat-singleturn`, `coding-singleturn`, `decode-heavy`, `osworld-multiturn-long`, `osworld-multiturn-medium`, `osworld-multiturn-short`, `prefill-heavy`, `random-1k`, `swebench-multiturn-medium`, `swebench-multiturn-short`, `terminalbench-multiturn-medium`, `terminalbench-multiturn-short`
320
 
321
+ ### synthetic_distributional (1,365 rows)
322
 
323
+ ISL/OSL sampled from lognormal fits to real workload statistics. 49 unique (model, hardware, engine) combinations, 20 profiles, 11 concurrency levels {1, 5, 10, 20, 40, 80, 120, 160, 200, 256, 320}, 6.5% matrix fill.
324
 
325
+ 20 profiles: `chat-medium`, `chat-multiturn`, `chat-multiturn-long`, `chat-multiturn-medium`, `chat-multiturn-short`, `chat-multiturn-synth`, `chat-short`, `chat-singleturn`, `chat-singleturn-synth`, `coding-singleturn`, `osworld-multiturn`, `osworld-multiturn-long`, `osworld-multiturn-medium`, `osworld-multiturn-short`, `osworld-multiturn-synth`, `swebench-multiturn`, `swebench-multiturn-synth`, `terminalbench-multiturn`, `terminalbench-multiturn-short`, `terminalbench-multiturn-synth`
326
 
327
  ### per_layer_kernel (37 rows)
328
 
 
338
 
339
  ### Quality filtering
340
 
341
+ Concurrency levels: trace_replay {1, 5, 10, 20, 40, 80}, synthetic_distributional {1, 5, 10, 20, 40, 80, 120, 160, 200, 256, 320}. Configurations where fewer than 75% of requests completed successfully are excluded. Summary metrics are computed from successful requests only.
342
 
343
  | Config | Rows |
344
  |--------|------|
345
  | trace_replay | 2,932 |
346
+ | synthetic_distributional | 1,365 |
347
  | per_layer_kernel | 37 |
348
  | kernels_labeled | 148,077 |
349
  | mse_validation | 28 |
croissant.json CHANGED
@@ -1,953 +1,953 @@
1
- {
2
- "@context": {
3
- "@language": "en",
4
- "@vocab": "https://schema.org/",
5
- "citeAs": "cr:citeAs",
6
- "column": "cr:column",
7
- "conformsTo": "dct:conformsTo",
8
- "cr": "http://mlcommons.org/croissant/",
9
- "data": {
10
- "@id": "cr:data",
11
- "@type": "@json"
12
- },
13
- "dataType": {
14
- "@id": "cr:dataType",
15
- "@type": "@vocab"
16
- },
17
- "dct": "http://purl.org/dc/terms/",
18
- "equivalentProperty": "cr:equivalentProperty",
19
- "examples": {
20
- "@id": "cr:examples",
21
- "@type": "@json"
22
- },
23
- "extract": "cr:extract",
24
- "field": "cr:field",
25
- "fileObject": "cr:fileObject",
26
- "fileProperty": "cr:fileProperty",
27
- "fileSet": "cr:fileSet",
28
- "format": "cr:format",
29
- "includes": "cr:includes",
30
- "isLiveDataset": "cr:isLiveDataset",
31
- "jsonPath": "cr:jsonPath",
32
- "key": "cr:key",
33
- "md5": "cr:md5",
34
- "parentField": "cr:parentField",
35
- "path": "cr:path",
36
- "rai": "http://mlcommons.org/croissant/RAI/",
37
- "recordSet": "cr:recordSet",
38
- "references": "cr:references",
39
- "regex": "cr:regex",
40
- "repeated": "cr:repeated",
41
- "replace": "cr:replace",
42
- "samplingRate": "cr:samplingRate",
43
- "prov": "http://www.w3.org/ns/prov#",
44
- "sc": "https://schema.org/",
45
- "separator": "cr:separator",
46
- "source": "cr:source",
47
- "subField": "cr:subField",
48
- "transform": "cr:transform"
49
- },
50
- "@type": "sc:Dataset",
51
- "name": "AgentPerfBench",
52
- "description": "LLM inference benchmark dataset: 3,155 main sweep rows measuring TTFT, TPOT, ITL, and throughput across 9 models, up to 14 GPU configurations, and 2 serving engines (vLLM 0.19.0, SGLang 0.5.9). Also includes 148,077 per-kernel NCU profiles, 28 MSE validation rows, and 37 per-layer kernel validation rows.",
53
- "url": "https://huggingface.co/datasets/agent-perf-bench/AgentPerfBench",
54
- "license": "https://spdx.org/licenses/Apache-2.0.html",
55
- "conformsTo": "http://mlcommons.org/croissant/1.1",
56
- "datePublished": "2026-05-04",
57
- "version": "1.0",
58
- "citeAs": "@inproceedings{agentperfbench2026, title={AgentPerfBench: A Benchmarking and Evaluation Suite for Inference Performance of Agentic LLMs}, author={Anonymous}, booktitle={NeurIPS 2026 Evaluations and Datasets Track}, year={2026}}",
59
- "creator": {
60
- "@type": "sc:Organization",
61
- "name": "Anonymous (under review)"
62
- },
63
- "distribution": [
64
- {
65
- "@type": "cr:FileObject",
66
- "@id": "trace-replay-summary-parquet",
67
- "name": "trace_replay/summary.parquet",
68
- "contentUrl": "https://huggingface.co/datasets/agent-perf-bench/AgentPerfBench/resolve/main/trace_replay/summary.parquet",
69
- "encodingFormat": "application/x-parquet",
70
- "sha256": "568efbdae473a830ba792317fb09cfba0ea4f667db67b93d28e72637a5c71d1e"
71
- },
72
- {
73
- "@type": "cr:FileObject",
74
- "@id": "synthetic-distributional-summary-parquet",
75
- "name": "synthetic_distributional/summary.parquet",
76
- "contentUrl": "https://huggingface.co/datasets/agent-perf-bench/AgentPerfBench/resolve/main/synthetic_distributional/summary.parquet",
77
- "encodingFormat": "application/x-parquet",
78
- "sha256": "0fd6ff9f802f9465a4618dfe8d5d426013028565e08ef5024f4302b9126a05a5"
79
- },
80
- {
81
- "@type": "cr:FileObject",
82
- "@id": "kernels-labeled-parquet",
83
- "name": "kernel_profiles/kernels_labeled.parquet",
84
- "contentUrl": "https://huggingface.co/datasets/agent-perf-bench/AgentPerfBench/resolve/main/kernel_profiles/kernels_labeled.parquet",
85
- "encodingFormat": "application/x-parquet",
86
- "sha256": "05571ee61fe97beac7f3d0b0e8be07a0feeca00cee9b308cd469a57cd4c9b671"
87
- },
88
- {
89
- "@type": "cr:FileObject",
90
- "@id": "mse-validation-summary-parquet",
91
- "name": "mse_validation/summary.parquet",
92
- "contentUrl": "https://huggingface.co/datasets/agent-perf-bench/AgentPerfBench/resolve/main/mse_validation/summary.parquet",
93
- "encodingFormat": "application/x-parquet",
94
- "sha256": "90cf7eec0e86473c92f0c36605000586992363902d612e3f25f4a58ac2811e4c"
95
- },
96
- {
97
- "@type": "cr:FileObject",
98
- "@id": "per-layer-kernel-summary-parquet",
99
- "name": "per_layer_kernel/summary.parquet",
100
- "contentUrl": "https://huggingface.co/datasets/agent-perf-bench/AgentPerfBench/resolve/main/per_layer_kernel/summary.parquet",
101
- "encodingFormat": "application/x-parquet",
102
- "sha256": "c78e4694009da435587bc9e7e3e4821fca6a3bf881284f873b71ccdb5ed47691"
103
- }
104
- ],
105
- "recordSet": [
106
- {
107
- "@type": "cr:RecordSet",
108
- "@id": "trace-replay-summary",
109
- "name": "Trace Replay Summary",
110
- "description": "One row per benchmark configuration from trace replay runs (empirical agent session replays).",
111
- "field": [
112
- {
113
- "@type": "cr:Field",
114
- "@id": "trace-replay-summary/run_id",
115
- "name": "run_id",
116
- "dataType": "sc:Text",
117
- "source": {
118
- "fileObject": {
119
- "@id": "trace-replay-summary-parquet"
120
- },
121
- "extract": {
122
- "column": "run_id"
123
- }
124
- }
125
- },
126
- {
127
- "@type": "cr:Field",
128
- "@id": "trace-replay-summary/model",
129
- "name": "model",
130
- "dataType": "sc:Text",
131
- "source": {
132
- "fileObject": {
133
- "@id": "trace-replay-summary-parquet"
134
- },
135
- "extract": {
136
- "column": "model"
137
- }
138
- }
139
- },
140
- {
141
- "@type": "cr:Field",
142
- "@id": "trace-replay-summary/hardware",
143
- "name": "hardware",
144
- "dataType": "sc:Text",
145
- "source": {
146
- "fileObject": {
147
- "@id": "trace-replay-summary-parquet"
148
- },
149
- "extract": {
150
- "column": "hardware"
151
- }
152
- }
153
- },
154
- {
155
- "@type": "cr:Field",
156
- "@id": "trace-replay-summary/engine",
157
- "name": "engine",
158
- "dataType": "sc:Text",
159
- "source": {
160
- "fileObject": {
161
- "@id": "trace-replay-summary-parquet"
162
- },
163
- "extract": {
164
- "column": "engine"
165
- }
166
- }
167
- },
168
- {
169
- "@type": "cr:Field",
170
- "@id": "trace-replay-summary/profile",
171
- "name": "profile",
172
- "dataType": "sc:Text",
173
- "source": {
174
- "fileObject": {
175
- "@id": "trace-replay-summary-parquet"
176
- },
177
- "extract": {
178
- "column": "profile"
179
- }
180
- }
181
- },
182
- {
183
- "@type": "cr:Field",
184
- "@id": "trace-replay-summary/concurrency",
185
- "name": "concurrency",
186
- "dataType": "sc:Integer",
187
- "source": {
188
- "fileObject": {
189
- "@id": "trace-replay-summary-parquet"
190
- },
191
- "extract": {
192
- "column": "concurrency"
193
- }
194
- }
195
- },
196
- {
197
- "@type": "cr:Field",
198
- "@id": "trace-replay-summary/request_throughput",
199
- "name": "request_throughput",
200
- "dataType": "sc:Float",
201
- "source": {
202
- "fileObject": {
203
- "@id": "trace-replay-summary-parquet"
204
- },
205
- "extract": {
206
- "column": "request_throughput"
207
- }
208
- }
209
- },
210
- {
211
- "@type": "cr:Field",
212
- "@id": "trace-replay-summary/median_ttft_ms",
213
- "name": "median_ttft_ms",
214
- "dataType": "sc:Float",
215
- "source": {
216
- "fileObject": {
217
- "@id": "trace-replay-summary-parquet"
218
- },
219
- "extract": {
220
- "column": "median_ttft_ms"
221
- }
222
- }
223
- },
224
- {
225
- "@type": "cr:Field",
226
- "@id": "trace-replay-summary/median_tpot_ms",
227
- "name": "median_tpot_ms",
228
- "dataType": "sc:Float",
229
- "source": {
230
- "fileObject": {
231
- "@id": "trace-replay-summary-parquet"
232
- },
233
- "extract": {
234
- "column": "median_tpot_ms"
235
- }
236
- }
237
- }
238
- ]
239
- },
240
- {
241
- "@type": "cr:RecordSet",
242
- "@id": "synthetic-distributional-summary",
243
- "name": "Synthetic Distributional Summary",
244
- "description": "One row per benchmark configuration from synthetic distributional runs (statistical sampling from lognormal fits to recorded workload statistics).",
245
- "field": [
246
- {
247
- "@type": "cr:Field",
248
- "@id": "synthetic-distributional-summary/run_id",
249
- "name": "run_id",
250
- "dataType": "sc:Text",
251
- "source": {
252
- "fileObject": {
253
- "@id": "synthetic-distributional-summary-parquet"
254
- },
255
- "extract": {
256
- "column": "run_id"
257
- }
258
- }
259
- },
260
- {
261
- "@type": "cr:Field",
262
- "@id": "synthetic-distributional-summary/model",
263
- "name": "model",
264
- "dataType": "sc:Text",
265
- "source": {
266
- "fileObject": {
267
- "@id": "synthetic-distributional-summary-parquet"
268
- },
269
- "extract": {
270
- "column": "model"
271
- }
272
- }
273
- },
274
- {
275
- "@type": "cr:Field",
276
- "@id": "synthetic-distributional-summary/hardware",
277
- "name": "hardware",
278
- "dataType": "sc:Text",
279
- "source": {
280
- "fileObject": {
281
- "@id": "synthetic-distributional-summary-parquet"
282
- },
283
- "extract": {
284
- "column": "hardware"
285
- }
286
- }
287
- },
288
- {
289
- "@type": "cr:Field",
290
- "@id": "synthetic-distributional-summary/engine",
291
- "name": "engine",
292
- "dataType": "sc:Text",
293
- "source": {
294
- "fileObject": {
295
- "@id": "synthetic-distributional-summary-parquet"
296
- },
297
- "extract": {
298
- "column": "engine"
299
- }
300
- }
301
- },
302
- {
303
- "@type": "cr:Field",
304
- "@id": "synthetic-distributional-summary/profile",
305
- "name": "profile",
306
- "dataType": "sc:Text",
307
- "source": {
308
- "fileObject": {
309
- "@id": "synthetic-distributional-summary-parquet"
310
- },
311
- "extract": {
312
- "column": "profile"
313
- }
314
- }
315
- },
316
- {
317
- "@type": "cr:Field",
318
- "@id": "synthetic-distributional-summary/concurrency",
319
- "name": "concurrency",
320
- "dataType": "sc:Integer",
321
- "source": {
322
- "fileObject": {
323
- "@id": "synthetic-distributional-summary-parquet"
324
- },
325
- "extract": {
326
- "column": "concurrency"
327
- }
328
- }
329
- },
330
- {
331
- "@type": "cr:Field",
332
- "@id": "synthetic-distributional-summary/request_throughput",
333
- "name": "request_throughput",
334
- "dataType": "sc:Float",
335
- "source": {
336
- "fileObject": {
337
- "@id": "synthetic-distributional-summary-parquet"
338
- },
339
- "extract": {
340
- "column": "request_throughput"
341
- }
342
- }
343
- },
344
- {
345
- "@type": "cr:Field",
346
- "@id": "synthetic-distributional-summary/median_ttft_ms",
347
- "name": "median_ttft_ms",
348
- "dataType": "sc:Float",
349
- "source": {
350
- "fileObject": {
351
- "@id": "synthetic-distributional-summary-parquet"
352
- },
353
- "extract": {
354
- "column": "median_ttft_ms"
355
- }
356
- }
357
- },
358
- {
359
- "@type": "cr:Field",
360
- "@id": "synthetic-distributional-summary/median_tpot_ms",
361
- "name": "median_tpot_ms",
362
- "dataType": "sc:Float",
363
- "source": {
364
- "fileObject": {
365
- "@id": "synthetic-distributional-summary-parquet"
366
- },
367
- "extract": {
368
- "column": "median_tpot_ms"
369
- }
370
- }
371
- }
372
- ]
373
- },
374
- {
375
- "@type": "cr:RecordSet",
376
- "@id": "kernels-labeled",
377
- "name": "Kernels Labeled",
378
- "description": "Per-kernel Nsight Compute (ncu) hardware counter profiles. Individual kernel invocations across 4 GPUs and 13 model/sweep sources.",
379
- "field": [
380
- {
381
- "@type": "cr:Field",
382
- "@id": "kernels-labeled/source",
383
- "name": "source",
384
- "dataType": "sc:Text",
385
- "source": {
386
- "fileObject": {
387
- "@id": "kernels-labeled-parquet"
388
- },
389
- "extract": {
390
- "column": "source"
391
- }
392
- }
393
- },
394
- {
395
- "@type": "cr:Field",
396
- "@id": "kernels-labeled/gpu",
397
- "name": "gpu",
398
- "dataType": "sc:Text",
399
- "source": {
400
- "fileObject": {
401
- "@id": "kernels-labeled-parquet"
402
- },
403
- "extract": {
404
- "column": "gpu"
405
- }
406
- }
407
- },
408
- {
409
- "@type": "cr:Field",
410
- "@id": "kernels-labeled/model",
411
- "name": "model",
412
- "dataType": "sc:Text",
413
- "source": {
414
- "fileObject": {
415
- "@id": "kernels-labeled-parquet"
416
- },
417
- "extract": {
418
- "column": "model"
419
- }
420
- }
421
- },
422
- {
423
- "@type": "cr:Field",
424
- "@id": "kernels-labeled/kernel_family",
425
- "name": "kernel_family",
426
- "dataType": "sc:Text",
427
- "source": {
428
- "fileObject": {
429
- "@id": "kernels-labeled-parquet"
430
- },
431
- "extract": {
432
- "column": "kernel_family"
433
- }
434
- }
435
- },
436
- {
437
- "@type": "cr:Field",
438
- "@id": "kernels-labeled/kernel_name",
439
- "name": "kernel_name",
440
- "dataType": "sc:Text",
441
- "source": {
442
- "fileObject": {
443
- "@id": "kernels-labeled-parquet"
444
- },
445
- "extract": {
446
- "column": "kernel_name"
447
- }
448
- }
449
- },
450
- {
451
- "@type": "cr:Field",
452
- "@id": "kernels-labeled/gpu_time_duration_ms",
453
- "name": "gpu_time_duration_ms",
454
- "dataType": "sc:Float",
455
- "source": {
456
- "fileObject": {
457
- "@id": "kernels-labeled-parquet"
458
- },
459
- "extract": {
460
- "column": "gpu_time_duration_ms"
461
- }
462
- }
463
- },
464
- {
465
- "@type": "cr:Field",
466
- "@id": "kernels-labeled/dram_bytes_sum",
467
- "name": "dram_bytes_sum",
468
- "dataType": "sc:Float",
469
- "source": {
470
- "fileObject": {
471
- "@id": "kernels-labeled-parquet"
472
- },
473
- "extract": {
474
- "column": "dram_bytes_sum"
475
- }
476
- }
477
- }
478
- ]
479
- },
480
- {
481
- "@type": "cr:RecordSet",
482
- "@id": "mse-validation-summary",
483
- "name": "MSE Validation Summary",
484
- "description": "Curated validation rows for the distributional synthetic replay generator. Paired synthetic and real trace replay runs on H100 / Llama-3.1-8B / vLLM.",
485
- "field": [
486
- {
487
- "@type": "cr:Field",
488
- "@id": "mse-validation-summary/validation_id",
489
- "name": "validation_id",
490
- "dataType": "sc:Text",
491
- "source": {
492
- "fileObject": {
493
- "@id": "mse-validation-summary-parquet"
494
- },
495
- "extract": {
496
- "column": "validation_id"
497
- }
498
- }
499
- },
500
- {
501
- "@type": "cr:Field",
502
- "@id": "mse-validation-summary/tier",
503
- "name": "tier",
504
- "dataType": "sc:Text",
505
- "source": {
506
- "fileObject": {
507
- "@id": "mse-validation-summary-parquet"
508
- },
509
- "extract": {
510
- "column": "tier"
511
- }
512
- }
513
- },
514
- {
515
- "@type": "cr:Field",
516
- "@id": "mse-validation-summary/stage",
517
- "name": "stage",
518
- "dataType": "sc:Text",
519
- "source": {
520
- "fileObject": {
521
- "@id": "mse-validation-summary-parquet"
522
- },
523
- "extract": {
524
- "column": "stage"
525
- }
526
- }
527
- },
528
- {
529
- "@type": "cr:Field",
530
- "@id": "mse-validation-summary/workload",
531
- "name": "workload",
532
- "dataType": "sc:Text",
533
- "source": {
534
- "fileObject": {
535
- "@id": "mse-validation-summary-parquet"
536
- },
537
- "extract": {
538
- "column": "workload"
539
- }
540
- }
541
- },
542
- {
543
- "@type": "cr:Field",
544
- "@id": "mse-validation-summary/profile",
545
- "name": "profile",
546
- "dataType": "sc:Text",
547
- "source": {
548
- "fileObject": {
549
- "@id": "mse-validation-summary-parquet"
550
- },
551
- "extract": {
552
- "column": "profile"
553
- }
554
- }
555
- },
556
- {
557
- "@type": "cr:Field",
558
- "@id": "mse-validation-summary/model",
559
- "name": "model",
560
- "dataType": "sc:Text",
561
- "source": {
562
- "fileObject": {
563
- "@id": "mse-validation-summary-parquet"
564
- },
565
- "extract": {
566
- "column": "model"
567
- }
568
- }
569
- },
570
- {
571
- "@type": "cr:Field",
572
- "@id": "mse-validation-summary/hardware",
573
- "name": "hardware",
574
- "dataType": "sc:Text",
575
- "source": {
576
- "fileObject": {
577
- "@id": "mse-validation-summary-parquet"
578
- },
579
- "extract": {
580
- "column": "hardware"
581
- }
582
- }
583
- },
584
- {
585
- "@type": "cr:Field",
586
- "@id": "mse-validation-summary/engine",
587
- "name": "engine",
588
- "dataType": "sc:Text",
589
- "source": {
590
- "fileObject": {
591
- "@id": "mse-validation-summary-parquet"
592
- },
593
- "extract": {
594
- "column": "engine"
595
- }
596
- }
597
- },
598
- {
599
- "@type": "cr:Field",
600
- "@id": "mse-validation-summary/concurrency",
601
- "name": "concurrency",
602
- "dataType": "sc:Integer",
603
- "source": {
604
- "fileObject": {
605
- "@id": "mse-validation-summary-parquet"
606
- },
607
- "extract": {
608
- "column": "concurrency"
609
- }
610
- }
611
- },
612
- {
613
- "@type": "cr:Field",
614
- "@id": "mse-validation-summary/success_rate",
615
- "name": "success_rate",
616
- "dataType": "sc:Float",
617
- "source": {
618
- "fileObject": {
619
- "@id": "mse-validation-summary-parquet"
620
- },
621
- "extract": {
622
- "column": "success_rate"
623
- }
624
- }
625
- },
626
- {
627
- "@type": "cr:Field",
628
- "@id": "mse-validation-summary/request_throughput",
629
- "name": "request_throughput",
630
- "dataType": "sc:Float",
631
- "source": {
632
- "fileObject": {
633
- "@id": "mse-validation-summary-parquet"
634
- },
635
- "extract": {
636
- "column": "request_throughput"
637
- }
638
- }
639
- },
640
- {
641
- "@type": "cr:Field",
642
- "@id": "mse-validation-summary/successful_requests",
643
- "name": "successful_requests",
644
- "dataType": "sc:Integer",
645
- "source": {
646
- "fileObject": {
647
- "@id": "mse-validation-summary-parquet"
648
- },
649
- "extract": {
650
- "column": "successful_requests"
651
- }
652
- }
653
- },
654
- {
655
- "@type": "cr:Field",
656
- "@id": "mse-validation-summary/failed_requests",
657
- "name": "failed_requests",
658
- "dataType": "sc:Integer",
659
- "source": {
660
- "fileObject": {
661
- "@id": "mse-validation-summary-parquet"
662
- },
663
- "extract": {
664
- "column": "failed_requests"
665
- }
666
- }
667
- },
668
- {
669
- "@type": "cr:Field",
670
- "@id": "mse-validation-summary/median_ttft_ms",
671
- "name": "median_ttft_ms",
672
- "dataType": "sc:Float",
673
- "source": {
674
- "fileObject": {
675
- "@id": "mse-validation-summary-parquet"
676
- },
677
- "extract": {
678
- "column": "median_ttft_ms"
679
- }
680
- }
681
- },
682
- {
683
- "@type": "cr:Field",
684
- "@id": "mse-validation-summary/median_tpot_ms",
685
- "name": "median_tpot_ms",
686
- "dataType": "sc:Float",
687
- "source": {
688
- "fileObject": {
689
- "@id": "mse-validation-summary-parquet"
690
- },
691
- "extract": {
692
- "column": "median_tpot_ms"
693
- }
694
- }
695
- }
696
- ]
697
- },
698
- {
699
- "@type": "cr:RecordSet",
700
- "@id": "per-layer-kernel-summary",
701
- "name": "Per-Layer Kernel Summary",
702
- "description": "Per-component analytical OI decomposition and NCU kernel profiles for Llama-3.1-8B on H100 (prefill phase).",
703
- "field": [
704
- {
705
- "@type": "cr:Field",
706
- "@id": "per-layer-kernel-summary/record_type",
707
- "name": "record_type",
708
- "dataType": "sc:Text",
709
- "source": {
710
- "fileObject": {
711
- "@id": "per-layer-kernel-summary-parquet"
712
- },
713
- "extract": {
714
- "column": "record_type"
715
- }
716
- }
717
- },
718
- {
719
- "@type": "cr:Field",
720
- "@id": "per-layer-kernel-summary/model",
721
- "name": "model",
722
- "dataType": "sc:Text",
723
- "source": {
724
- "fileObject": {
725
- "@id": "per-layer-kernel-summary-parquet"
726
- },
727
- "extract": {
728
- "column": "model"
729
- }
730
- }
731
- },
732
- {
733
- "@type": "cr:Field",
734
- "@id": "per-layer-kernel-summary/hardware",
735
- "name": "hardware",
736
- "dataType": "sc:Text",
737
- "source": {
738
- "fileObject": {
739
- "@id": "per-layer-kernel-summary-parquet"
740
- },
741
- "extract": {
742
- "column": "hardware"
743
- }
744
- }
745
- },
746
- {
747
- "@type": "cr:Field",
748
- "@id": "per-layer-kernel-summary/phase",
749
- "name": "phase",
750
- "dataType": "sc:Text",
751
- "source": {
752
- "fileObject": {
753
- "@id": "per-layer-kernel-summary-parquet"
754
- },
755
- "extract": {
756
- "column": "phase"
757
- }
758
- }
759
- },
760
- {
761
- "@type": "cr:Field",
762
- "@id": "per-layer-kernel-summary/batch_size",
763
- "name": "batch_size",
764
- "dataType": "sc:Integer",
765
- "source": {
766
- "fileObject": {
767
- "@id": "per-layer-kernel-summary-parquet"
768
- },
769
- "extract": {
770
- "column": "batch_size"
771
- }
772
- }
773
- },
774
- {
775
- "@type": "cr:Field",
776
- "@id": "per-layer-kernel-summary/sequence_length",
777
- "name": "sequence_length",
778
- "dataType": "sc:Integer",
779
- "source": {
780
- "fileObject": {
781
- "@id": "per-layer-kernel-summary-parquet"
782
- },
783
- "extract": {
784
- "column": "sequence_length"
785
- }
786
- }
787
- },
788
- {
789
- "@type": "cr:Field",
790
- "@id": "per-layer-kernel-summary/component_name",
791
- "name": "component_name",
792
- "dataType": "sc:Text",
793
- "source": {
794
- "fileObject": {
795
- "@id": "per-layer-kernel-summary-parquet"
796
- },
797
- "extract": {
798
- "column": "component_name"
799
- }
800
- }
801
- },
802
- {
803
- "@type": "cr:Field",
804
- "@id": "per-layer-kernel-summary/bound",
805
- "name": "bound",
806
- "dataType": "sc:Text",
807
- "source": {
808
- "fileObject": {
809
- "@id": "per-layer-kernel-summary-parquet"
810
- },
811
- "extract": {
812
- "column": "bound"
813
- }
814
- }
815
- },
816
- {
817
- "@type": "cr:Field",
818
- "@id": "per-layer-kernel-summary/flops",
819
- "name": "flops",
820
- "dataType": "sc:Float",
821
- "source": {
822
- "fileObject": {
823
- "@id": "per-layer-kernel-summary-parquet"
824
- },
825
- "extract": {
826
- "column": "flops"
827
- }
828
- }
829
- },
830
- {
831
- "@type": "cr:Field",
832
- "@id": "per-layer-kernel-summary/bytes_accessed",
833
- "name": "bytes_accessed",
834
- "dataType": "sc:Float",
835
- "source": {
836
- "fileObject": {
837
- "@id": "per-layer-kernel-summary-parquet"
838
- },
839
- "extract": {
840
- "column": "bytes_accessed"
841
- }
842
- }
843
- },
844
- {
845
- "@type": "cr:Field",
846
- "@id": "per-layer-kernel-summary/operational_intensity",
847
- "name": "operational_intensity",
848
- "dataType": "sc:Float",
849
- "source": {
850
- "fileObject": {
851
- "@id": "per-layer-kernel-summary-parquet"
852
- },
853
- "extract": {
854
- "column": "operational_intensity"
855
- }
856
- }
857
- },
858
- {
859
- "@type": "cr:Field",
860
- "@id": "per-layer-kernel-summary/kernel_name",
861
- "name": "kernel_name",
862
- "dataType": "sc:Text",
863
- "source": {
864
- "fileObject": {
865
- "@id": "per-layer-kernel-summary-parquet"
866
- },
867
- "extract": {
868
- "column": "kernel_name"
869
- }
870
- }
871
- },
872
- {
873
- "@type": "cr:Field",
874
- "@id": "per-layer-kernel-summary/duration_us",
875
- "name": "duration_us",
876
- "dataType": "sc:Float",
877
- "source": {
878
- "fileObject": {
879
- "@id": "per-layer-kernel-summary-parquet"
880
- },
881
- "extract": {
882
- "column": "duration_us"
883
- }
884
- }
885
- },
886
- {
887
- "@type": "cr:Field",
888
- "@id": "per-layer-kernel-summary/compute_sm_throughput_pct",
889
- "name": "compute_sm_throughput_pct",
890
- "dataType": "sc:Float",
891
- "source": {
892
- "fileObject": {
893
- "@id": "per-layer-kernel-summary-parquet"
894
- },
895
- "extract": {
896
- "column": "compute_sm_throughput_pct"
897
- }
898
- }
899
- },
900
- {
901
- "@type": "cr:Field",
902
- "@id": "per-layer-kernel-summary/dram_throughput_pct",
903
- "name": "dram_throughput_pct",
904
- "dataType": "sc:Float",
905
- "source": {
906
- "fileObject": {
907
- "@id": "per-layer-kernel-summary-parquet"
908
- },
909
- "extract": {
910
- "column": "dram_throughput_pct"
911
- }
912
- }
913
- }
914
- ]
915
- }
916
- ],
917
- "rai:dataLimitations": "Results cover NVIDIA H100, A100, RTX 3090, and RTX 2080 Ti GPUs only and may not generalize to other accelerators (AMD, Intel, TPU). Benchmark configurations are pinned to vLLM 0.19.0 and SGLang 0.5.9; results do not represent other engine versions. Concurrency levels (1-320) may not cover extreme-scale deployments. Not recommended as sole basis for hardware purchasing decisions or for comparing model task quality.",
918
- "rai:dataBiases": "Model selection over-represents Meta Llama and Alibaba Qwen families. Hardware is exclusively NVIDIA GPUs. Workload profiles are author-designed approximations of production traffic; real deployment patterns may differ.",
919
- "rai:personalSensitiveInformation": "No personally identifiable information is present. All API endpoints and credentials are stripped. Workload traces use synthetic random tokens or publicly available coding benchmarks.",
920
- "rai:dataUseCases": "Established uses: relative comparison of inference engine throughput, latency benchmarking under controlled conditions, TTFT scaling with context length in multi-turn sessions, per-kernel performance modelling. Not established: absolute latency prediction for production, model quality comparison, cost estimation.",
921
- "rai:dataSocialImpact": "Enables reproducible comparison of open-source LLM serving systems, supporting infrastructure research and reducing vendor lock-in.",
922
- "rai:hasSyntheticData": true,
923
- "prov:wasDerivedFrom": [
924
- {
925
- "@id": "https://huggingface.co/datasets/princeton-nlp/SWE-bench",
926
- "name": "SWE-Bench"
927
- },
928
- {
929
- "@id": "https://github.com/TerminalBench/TerminalBench",
930
- "name": "TerminalBench"
931
- },
932
- {
933
- "@id": "https://github.com/xlang-ai/OSWorld",
934
- "name": "OSWorld"
935
- },
936
- {
937
- "@id": "https://huggingface.co/datasets/Aeala/ShareGPT_Vicuna_unfiltered",
938
- "name": "ShareGPT"
939
- }
940
- ],
941
- "prov:wasGeneratedBy": {
942
- "@type": "prov:Activity",
943
- "name": "AgentPerfBench benchmark collection",
944
- "description": "Deploy model on target GPU with specified engine and tensor parallelism. Send requests per configuration after warmup using closed-loop concurrency control. Record per-request TTFT, TPOT, ITL, E2EL, and token counts. Compute summary statistics. Sanitize credentials and convert to Parquet."
945
- },
946
- "rai:dataCollection": "Automated benchmark collection. Models deployed on target GPU with specified engine and tensor parallelism. Requests sent per workload profile using closed-loop concurrency control with semaphore. Per-request TTFT, TPOT, ITL, E2EL, and token counts recorded. Kernel profiles collected via Nsight Compute (ncu) with hardware counter metrics.",
947
- "rai:dataCollectionType": "Automated measurement via benchmarking scripts. No human subjects or crowdsourcing.",
948
- "rai:dataCollectionMissingData": "Some model-hardware-concurrency combinations are absent due to OOM or engine incompatibility. Concurrency levels exceeding the session-pool size were filtered (see Concurrency filtering in README).",
949
- "rai:dataPreprocessing": "Credentials and API endpoints stripped. Configurations with fewer than 75% of requests completing successfully are excluded. Summary statistics computed from successful requests. Kernel names demangled from NCU output.",
950
- "rai:dataAnnotationProtocol": "Not applicable. Fully automated benchmark collection with no human annotation.",
951
- "rai:dataAnnotationPlatform": "Not applicable. No human annotation.",
952
- "rai:dataAnnotationAnalysis": "Not applicable. No human annotation."
953
- }
 
1
+ {
2
+ "@context": {
3
+ "@language": "en",
4
+ "@vocab": "https://schema.org/",
5
+ "citeAs": "cr:citeAs",
6
+ "column": "cr:column",
7
+ "conformsTo": "dct:conformsTo",
8
+ "cr": "http://mlcommons.org/croissant/",
9
+ "data": {
10
+ "@id": "cr:data",
11
+ "@type": "@json"
12
+ },
13
+ "dataType": {
14
+ "@id": "cr:dataType",
15
+ "@type": "@vocab"
16
+ },
17
+ "dct": "http://purl.org/dc/terms/",
18
+ "equivalentProperty": "cr:equivalentProperty",
19
+ "examples": {
20
+ "@id": "cr:examples",
21
+ "@type": "@json"
22
+ },
23
+ "extract": "cr:extract",
24
+ "field": "cr:field",
25
+ "fileObject": "cr:fileObject",
26
+ "fileProperty": "cr:fileProperty",
27
+ "fileSet": "cr:fileSet",
28
+ "format": "cr:format",
29
+ "includes": "cr:includes",
30
+ "isLiveDataset": "cr:isLiveDataset",
31
+ "jsonPath": "cr:jsonPath",
32
+ "key": "cr:key",
33
+ "md5": "cr:md5",
34
+ "parentField": "cr:parentField",
35
+ "path": "cr:path",
36
+ "rai": "http://mlcommons.org/croissant/RAI/",
37
+ "recordSet": "cr:recordSet",
38
+ "references": "cr:references",
39
+ "regex": "cr:regex",
40
+ "repeated": "cr:repeated",
41
+ "replace": "cr:replace",
42
+ "samplingRate": "cr:samplingRate",
43
+ "prov": "http://www.w3.org/ns/prov#",
44
+ "sc": "https://schema.org/",
45
+ "separator": "cr:separator",
46
+ "source": "cr:source",
47
+ "subField": "cr:subField",
48
+ "transform": "cr:transform"
49
+ },
50
+ "@type": "sc:Dataset",
51
+ "name": "AgentPerfBench",
52
+ "description": "LLM inference benchmark dataset: 4,297 main sweep rows measuring TTFT, TPOT, ITL, and throughput across 9 models, up to 14 GPU configurations, and 2 serving engines (vLLM 0.19.0, SGLang 0.5.9). Also includes 148,077 per-kernel NCU profiles, 28 MSE validation rows, and 37 per-layer kernel validation rows.",
53
+ "url": "https://huggingface.co/datasets/agent-perf-bench/AgentPerfBench",
54
+ "license": "https://spdx.org/licenses/Apache-2.0.html",
55
+ "conformsTo": "http://mlcommons.org/croissant/1.1",
56
+ "datePublished": "2026-05-04",
57
+ "version": "1.0",
58
+ "citeAs": "@inproceedings{agentperfbench2026, title={AgentPerfBench: A Benchmarking and Evaluation Suite for Inference Performance of Agentic LLMs}, author={Anonymous}, booktitle={NeurIPS 2026 Evaluations and Datasets Track}, year={2026}}",
59
+ "creator": {
60
+ "@type": "sc:Organization",
61
+ "name": "Anonymous (under review)"
62
+ },
63
+ "distribution": [
64
+ {
65
+ "@type": "cr:FileObject",
66
+ "@id": "trace-replay-summary-parquet",
67
+ "name": "trace_replay/summary.parquet",
68
+ "contentUrl": "https://huggingface.co/datasets/agent-perf-bench/AgentPerfBench/resolve/main/trace_replay/summary.parquet",
69
+ "encodingFormat": "application/x-parquet",
70
+ "sha256": "568efbdae473a830ba792317fb09cfba0ea4f667db67b93d28e72637a5c71d1e"
71
+ },
72
+ {
73
+ "@type": "cr:FileObject",
74
+ "@id": "synthetic-distributional-summary-parquet",
75
+ "name": "synthetic_distributional/summary.parquet",
76
+ "contentUrl": "https://huggingface.co/datasets/agent-perf-bench/AgentPerfBench/resolve/main/synthetic_distributional/summary.parquet",
77
+ "encodingFormat": "application/x-parquet",
78
+ "sha256": "7e641197686e02248ff063fceb9046efaf03e5e2e7fbb50d5d57f785e17c9157"
79
+ },
80
+ {
81
+ "@type": "cr:FileObject",
82
+ "@id": "kernels-labeled-parquet",
83
+ "name": "kernel_profiles/kernels_labeled.parquet",
84
+ "contentUrl": "https://huggingface.co/datasets/agent-perf-bench/AgentPerfBench/resolve/main/kernel_profiles/kernels_labeled.parquet",
85
+ "encodingFormat": "application/x-parquet",
86
+ "sha256": "f9822c8c872bbfc00bc69d335ebef23bb8603f9a9e624267c56905f1e8cf8aa8"
87
+ },
88
+ {
89
+ "@type": "cr:FileObject",
90
+ "@id": "mse-validation-summary-parquet",
91
+ "name": "mse_validation/summary.parquet",
92
+ "contentUrl": "https://huggingface.co/datasets/agent-perf-bench/AgentPerfBench/resolve/main/mse_validation/summary.parquet",
93
+ "encodingFormat": "application/x-parquet",
94
+ "sha256": "90cf7eec0e86473c92f0c36605000586992363902d612e3f25f4a58ac2811e4c"
95
+ },
96
+ {
97
+ "@type": "cr:FileObject",
98
+ "@id": "per-layer-kernel-summary-parquet",
99
+ "name": "per_layer_kernel/summary.parquet",
100
+ "contentUrl": "https://huggingface.co/datasets/agent-perf-bench/AgentPerfBench/resolve/main/per_layer_kernel/summary.parquet",
101
+ "encodingFormat": "application/x-parquet",
102
+ "sha256": "c78e4694009da435587bc9e7e3e4821fca6a3bf881284f873b71ccdb5ed47691"
103
+ }
104
+ ],
105
+ "recordSet": [
106
+ {
107
+ "@type": "cr:RecordSet",
108
+ "@id": "trace-replay-summary",
109
+ "name": "Trace Replay Summary",
110
+ "description": "One row per benchmark configuration from trace replay runs (empirical agent session replays).",
111
+ "field": [
112
+ {
113
+ "@type": "cr:Field",
114
+ "@id": "trace-replay-summary/run_id",
115
+ "name": "run_id",
116
+ "dataType": "sc:Text",
117
+ "source": {
118
+ "fileObject": {
119
+ "@id": "trace-replay-summary-parquet"
120
+ },
121
+ "extract": {
122
+ "column": "run_id"
123
+ }
124
+ }
125
+ },
126
+ {
127
+ "@type": "cr:Field",
128
+ "@id": "trace-replay-summary/model",
129
+ "name": "model",
130
+ "dataType": "sc:Text",
131
+ "source": {
132
+ "fileObject": {
133
+ "@id": "trace-replay-summary-parquet"
134
+ },
135
+ "extract": {
136
+ "column": "model"
137
+ }
138
+ }
139
+ },
140
+ {
141
+ "@type": "cr:Field",
142
+ "@id": "trace-replay-summary/hardware",
143
+ "name": "hardware",
144
+ "dataType": "sc:Text",
145
+ "source": {
146
+ "fileObject": {
147
+ "@id": "trace-replay-summary-parquet"
148
+ },
149
+ "extract": {
150
+ "column": "hardware"
151
+ }
152
+ }
153
+ },
154
+ {
155
+ "@type": "cr:Field",
156
+ "@id": "trace-replay-summary/engine",
157
+ "name": "engine",
158
+ "dataType": "sc:Text",
159
+ "source": {
160
+ "fileObject": {
161
+ "@id": "trace-replay-summary-parquet"
162
+ },
163
+ "extract": {
164
+ "column": "engine"
165
+ }
166
+ }
167
+ },
168
+ {
169
+ "@type": "cr:Field",
170
+ "@id": "trace-replay-summary/profile",
171
+ "name": "profile",
172
+ "dataType": "sc:Text",
173
+ "source": {
174
+ "fileObject": {
175
+ "@id": "trace-replay-summary-parquet"
176
+ },
177
+ "extract": {
178
+ "column": "profile"
179
+ }
180
+ }
181
+ },
182
+ {
183
+ "@type": "cr:Field",
184
+ "@id": "trace-replay-summary/concurrency",
185
+ "name": "concurrency",
186
+ "dataType": "sc:Integer",
187
+ "source": {
188
+ "fileObject": {
189
+ "@id": "trace-replay-summary-parquet"
190
+ },
191
+ "extract": {
192
+ "column": "concurrency"
193
+ }
194
+ }
195
+ },
196
+ {
197
+ "@type": "cr:Field",
198
+ "@id": "trace-replay-summary/request_throughput",
199
+ "name": "request_throughput",
200
+ "dataType": "sc:Float",
201
+ "source": {
202
+ "fileObject": {
203
+ "@id": "trace-replay-summary-parquet"
204
+ },
205
+ "extract": {
206
+ "column": "request_throughput"
207
+ }
208
+ }
209
+ },
210
+ {
211
+ "@type": "cr:Field",
212
+ "@id": "trace-replay-summary/median_ttft_ms",
213
+ "name": "median_ttft_ms",
214
+ "dataType": "sc:Float",
215
+ "source": {
216
+ "fileObject": {
217
+ "@id": "trace-replay-summary-parquet"
218
+ },
219
+ "extract": {
220
+ "column": "median_ttft_ms"
221
+ }
222
+ }
223
+ },
224
+ {
225
+ "@type": "cr:Field",
226
+ "@id": "trace-replay-summary/median_tpot_ms",
227
+ "name": "median_tpot_ms",
228
+ "dataType": "sc:Float",
229
+ "source": {
230
+ "fileObject": {
231
+ "@id": "trace-replay-summary-parquet"
232
+ },
233
+ "extract": {
234
+ "column": "median_tpot_ms"
235
+ }
236
+ }
237
+ }
238
+ ]
239
+ },
240
+ {
241
+ "@type": "cr:RecordSet",
242
+ "@id": "synthetic-distributional-summary",
243
+ "name": "Synthetic Distributional Summary",
244
+ "description": "One row per benchmark configuration from synthetic distributional runs (statistical sampling from lognormal fits to recorded workload statistics).",
245
+ "field": [
246
+ {
247
+ "@type": "cr:Field",
248
+ "@id": "synthetic-distributional-summary/run_id",
249
+ "name": "run_id",
250
+ "dataType": "sc:Text",
251
+ "source": {
252
+ "fileObject": {
253
+ "@id": "synthetic-distributional-summary-parquet"
254
+ },
255
+ "extract": {
256
+ "column": "run_id"
257
+ }
258
+ }
259
+ },
260
+ {
261
+ "@type": "cr:Field",
262
+ "@id": "synthetic-distributional-summary/model",
263
+ "name": "model",
264
+ "dataType": "sc:Text",
265
+ "source": {
266
+ "fileObject": {
267
+ "@id": "synthetic-distributional-summary-parquet"
268
+ },
269
+ "extract": {
270
+ "column": "model"
271
+ }
272
+ }
273
+ },
274
+ {
275
+ "@type": "cr:Field",
276
+ "@id": "synthetic-distributional-summary/hardware",
277
+ "name": "hardware",
278
+ "dataType": "sc:Text",
279
+ "source": {
280
+ "fileObject": {
281
+ "@id": "synthetic-distributional-summary-parquet"
282
+ },
283
+ "extract": {
284
+ "column": "hardware"
285
+ }
286
+ }
287
+ },
288
+ {
289
+ "@type": "cr:Field",
290
+ "@id": "synthetic-distributional-summary/engine",
291
+ "name": "engine",
292
+ "dataType": "sc:Text",
293
+ "source": {
294
+ "fileObject": {
295
+ "@id": "synthetic-distributional-summary-parquet"
296
+ },
297
+ "extract": {
298
+ "column": "engine"
299
+ }
300
+ }
301
+ },
302
+ {
303
+ "@type": "cr:Field",
304
+ "@id": "synthetic-distributional-summary/profile",
305
+ "name": "profile",
306
+ "dataType": "sc:Text",
307
+ "source": {
308
+ "fileObject": {
309
+ "@id": "synthetic-distributional-summary-parquet"
310
+ },
311
+ "extract": {
312
+ "column": "profile"
313
+ }
314
+ }
315
+ },
316
+ {
317
+ "@type": "cr:Field",
318
+ "@id": "synthetic-distributional-summary/concurrency",
319
+ "name": "concurrency",
320
+ "dataType": "sc:Integer",
321
+ "source": {
322
+ "fileObject": {
323
+ "@id": "synthetic-distributional-summary-parquet"
324
+ },
325
+ "extract": {
326
+ "column": "concurrency"
327
+ }
328
+ }
329
+ },
330
+ {
331
+ "@type": "cr:Field",
332
+ "@id": "synthetic-distributional-summary/request_throughput",
333
+ "name": "request_throughput",
334
+ "dataType": "sc:Float",
335
+ "source": {
336
+ "fileObject": {
337
+ "@id": "synthetic-distributional-summary-parquet"
338
+ },
339
+ "extract": {
340
+ "column": "request_throughput"
341
+ }
342
+ }
343
+ },
344
+ {
345
+ "@type": "cr:Field",
346
+ "@id": "synthetic-distributional-summary/median_ttft_ms",
347
+ "name": "median_ttft_ms",
348
+ "dataType": "sc:Float",
349
+ "source": {
350
+ "fileObject": {
351
+ "@id": "synthetic-distributional-summary-parquet"
352
+ },
353
+ "extract": {
354
+ "column": "median_ttft_ms"
355
+ }
356
+ }
357
+ },
358
+ {
359
+ "@type": "cr:Field",
360
+ "@id": "synthetic-distributional-summary/median_tpot_ms",
361
+ "name": "median_tpot_ms",
362
+ "dataType": "sc:Float",
363
+ "source": {
364
+ "fileObject": {
365
+ "@id": "synthetic-distributional-summary-parquet"
366
+ },
367
+ "extract": {
368
+ "column": "median_tpot_ms"
369
+ }
370
+ }
371
+ }
372
+ ]
373
+ },
374
+ {
375
+ "@type": "cr:RecordSet",
376
+ "@id": "kernels-labeled",
377
+ "name": "Kernels Labeled",
378
+ "description": "Per-kernel Nsight Compute (ncu) hardware counter profiles. Individual kernel invocations across 4 GPUs and 13 model/sweep sources.",
379
+ "field": [
380
+ {
381
+ "@type": "cr:Field",
382
+ "@id": "kernels-labeled/source",
383
+ "name": "source",
384
+ "dataType": "sc:Text",
385
+ "source": {
386
+ "fileObject": {
387
+ "@id": "kernels-labeled-parquet"
388
+ },
389
+ "extract": {
390
+ "column": "source"
391
+ }
392
+ }
393
+ },
394
+ {
395
+ "@type": "cr:Field",
396
+ "@id": "kernels-labeled/gpu",
397
+ "name": "gpu",
398
+ "dataType": "sc:Text",
399
+ "source": {
400
+ "fileObject": {
401
+ "@id": "kernels-labeled-parquet"
402
+ },
403
+ "extract": {
404
+ "column": "gpu"
405
+ }
406
+ }
407
+ },
408
+ {
409
+ "@type": "cr:Field",
410
+ "@id": "kernels-labeled/model",
411
+ "name": "model",
412
+ "dataType": "sc:Text",
413
+ "source": {
414
+ "fileObject": {
415
+ "@id": "kernels-labeled-parquet"
416
+ },
417
+ "extract": {
418
+ "column": "model"
419
+ }
420
+ }
421
+ },
422
+ {
423
+ "@type": "cr:Field",
424
+ "@id": "kernels-labeled/kernel_family",
425
+ "name": "kernel_family",
426
+ "dataType": "sc:Text",
427
+ "source": {
428
+ "fileObject": {
429
+ "@id": "kernels-labeled-parquet"
430
+ },
431
+ "extract": {
432
+ "column": "kernel_family"
433
+ }
434
+ }
435
+ },
436
+ {
437
+ "@type": "cr:Field",
438
+ "@id": "kernels-labeled/kernel_name",
439
+ "name": "kernel_name",
440
+ "dataType": "sc:Text",
441
+ "source": {
442
+ "fileObject": {
443
+ "@id": "kernels-labeled-parquet"
444
+ },
445
+ "extract": {
446
+ "column": "kernel_name"
447
+ }
448
+ }
449
+ },
450
+ {
451
+ "@type": "cr:Field",
452
+ "@id": "kernels-labeled/gpu_time_duration_ms",
453
+ "name": "gpu_time_duration_ms",
454
+ "dataType": "sc:Float",
455
+ "source": {
456
+ "fileObject": {
457
+ "@id": "kernels-labeled-parquet"
458
+ },
459
+ "extract": {
460
+ "column": "gpu_time_duration_ms"
461
+ }
462
+ }
463
+ },
464
+ {
465
+ "@type": "cr:Field",
466
+ "@id": "kernels-labeled/dram_bytes_sum",
467
+ "name": "dram_bytes_sum",
468
+ "dataType": "sc:Float",
469
+ "source": {
470
+ "fileObject": {
471
+ "@id": "kernels-labeled-parquet"
472
+ },
473
+ "extract": {
474
+ "column": "dram_bytes_sum"
475
+ }
476
+ }
477
+ }
478
+ ]
479
+ },
480
+ {
481
+ "@type": "cr:RecordSet",
482
+ "@id": "mse-validation-summary",
483
+ "name": "MSE Validation Summary",
484
+ "description": "Curated validation rows for the distributional synthetic replay generator. Paired synthetic and real trace replay runs on H100 / Llama-3.1-8B / vLLM.",
485
+ "field": [
486
+ {
487
+ "@type": "cr:Field",
488
+ "@id": "mse-validation-summary/validation_id",
489
+ "name": "validation_id",
490
+ "dataType": "sc:Text",
491
+ "source": {
492
+ "fileObject": {
493
+ "@id": "mse-validation-summary-parquet"
494
+ },
495
+ "extract": {
496
+ "column": "validation_id"
497
+ }
498
+ }
499
+ },
500
+ {
501
+ "@type": "cr:Field",
502
+ "@id": "mse-validation-summary/tier",
503
+ "name": "tier",
504
+ "dataType": "sc:Text",
505
+ "source": {
506
+ "fileObject": {
507
+ "@id": "mse-validation-summary-parquet"
508
+ },
509
+ "extract": {
510
+ "column": "tier"
511
+ }
512
+ }
513
+ },
514
+ {
515
+ "@type": "cr:Field",
516
+ "@id": "mse-validation-summary/stage",
517
+ "name": "stage",
518
+ "dataType": "sc:Text",
519
+ "source": {
520
+ "fileObject": {
521
+ "@id": "mse-validation-summary-parquet"
522
+ },
523
+ "extract": {
524
+ "column": "stage"
525
+ }
526
+ }
527
+ },
528
+ {
529
+ "@type": "cr:Field",
530
+ "@id": "mse-validation-summary/workload",
531
+ "name": "workload",
532
+ "dataType": "sc:Text",
533
+ "source": {
534
+ "fileObject": {
535
+ "@id": "mse-validation-summary-parquet"
536
+ },
537
+ "extract": {
538
+ "column": "workload"
539
+ }
540
+ }
541
+ },
542
+ {
543
+ "@type": "cr:Field",
544
+ "@id": "mse-validation-summary/profile",
545
+ "name": "profile",
546
+ "dataType": "sc:Text",
547
+ "source": {
548
+ "fileObject": {
549
+ "@id": "mse-validation-summary-parquet"
550
+ },
551
+ "extract": {
552
+ "column": "profile"
553
+ }
554
+ }
555
+ },
556
+ {
557
+ "@type": "cr:Field",
558
+ "@id": "mse-validation-summary/model",
559
+ "name": "model",
560
+ "dataType": "sc:Text",
561
+ "source": {
562
+ "fileObject": {
563
+ "@id": "mse-validation-summary-parquet"
564
+ },
565
+ "extract": {
566
+ "column": "model"
567
+ }
568
+ }
569
+ },
570
+ {
571
+ "@type": "cr:Field",
572
+ "@id": "mse-validation-summary/hardware",
573
+ "name": "hardware",
574
+ "dataType": "sc:Text",
575
+ "source": {
576
+ "fileObject": {
577
+ "@id": "mse-validation-summary-parquet"
578
+ },
579
+ "extract": {
580
+ "column": "hardware"
581
+ }
582
+ }
583
+ },
584
+ {
585
+ "@type": "cr:Field",
586
+ "@id": "mse-validation-summary/engine",
587
+ "name": "engine",
588
+ "dataType": "sc:Text",
589
+ "source": {
590
+ "fileObject": {
591
+ "@id": "mse-validation-summary-parquet"
592
+ },
593
+ "extract": {
594
+ "column": "engine"
595
+ }
596
+ }
597
+ },
598
+ {
599
+ "@type": "cr:Field",
600
+ "@id": "mse-validation-summary/concurrency",
601
+ "name": "concurrency",
602
+ "dataType": "sc:Integer",
603
+ "source": {
604
+ "fileObject": {
605
+ "@id": "mse-validation-summary-parquet"
606
+ },
607
+ "extract": {
608
+ "column": "concurrency"
609
+ }
610
+ }
611
+ },
612
+ {
613
+ "@type": "cr:Field",
614
+ "@id": "mse-validation-summary/success_rate",
615
+ "name": "success_rate",
616
+ "dataType": "sc:Float",
617
+ "source": {
618
+ "fileObject": {
619
+ "@id": "mse-validation-summary-parquet"
620
+ },
621
+ "extract": {
622
+ "column": "success_rate"
623
+ }
624
+ }
625
+ },
626
+ {
627
+ "@type": "cr:Field",
628
+ "@id": "mse-validation-summary/request_throughput",
629
+ "name": "request_throughput",
630
+ "dataType": "sc:Float",
631
+ "source": {
632
+ "fileObject": {
633
+ "@id": "mse-validation-summary-parquet"
634
+ },
635
+ "extract": {
636
+ "column": "request_throughput"
637
+ }
638
+ }
639
+ },
640
+ {
641
+ "@type": "cr:Field",
642
+ "@id": "mse-validation-summary/successful_requests",
643
+ "name": "successful_requests",
644
+ "dataType": "sc:Integer",
645
+ "source": {
646
+ "fileObject": {
647
+ "@id": "mse-validation-summary-parquet"
648
+ },
649
+ "extract": {
650
+ "column": "successful_requests"
651
+ }
652
+ }
653
+ },
654
+ {
655
+ "@type": "cr:Field",
656
+ "@id": "mse-validation-summary/failed_requests",
657
+ "name": "failed_requests",
658
+ "dataType": "sc:Integer",
659
+ "source": {
660
+ "fileObject": {
661
+ "@id": "mse-validation-summary-parquet"
662
+ },
663
+ "extract": {
664
+ "column": "failed_requests"
665
+ }
666
+ }
667
+ },
668
+ {
669
+ "@type": "cr:Field",
670
+ "@id": "mse-validation-summary/median_ttft_ms",
671
+ "name": "median_ttft_ms",
672
+ "dataType": "sc:Float",
673
+ "source": {
674
+ "fileObject": {
675
+ "@id": "mse-validation-summary-parquet"
676
+ },
677
+ "extract": {
678
+ "column": "median_ttft_ms"
679
+ }
680
+ }
681
+ },
682
+ {
683
+ "@type": "cr:Field",
684
+ "@id": "mse-validation-summary/median_tpot_ms",
685
+ "name": "median_tpot_ms",
686
+ "dataType": "sc:Float",
687
+ "source": {
688
+ "fileObject": {
689
+ "@id": "mse-validation-summary-parquet"
690
+ },
691
+ "extract": {
692
+ "column": "median_tpot_ms"
693
+ }
694
+ }
695
+ }
696
+ ]
697
+ },
698
+ {
699
+ "@type": "cr:RecordSet",
700
+ "@id": "per-layer-kernel-summary",
701
+ "name": "Per-Layer Kernel Summary",
702
+ "description": "Per-component analytical OI decomposition and NCU kernel profiles for Llama-3.1-8B on H100 (prefill phase).",
703
+ "field": [
704
+ {
705
+ "@type": "cr:Field",
706
+ "@id": "per-layer-kernel-summary/record_type",
707
+ "name": "record_type",
708
+ "dataType": "sc:Text",
709
+ "source": {
710
+ "fileObject": {
711
+ "@id": "per-layer-kernel-summary-parquet"
712
+ },
713
+ "extract": {
714
+ "column": "record_type"
715
+ }
716
+ }
717
+ },
718
+ {
719
+ "@type": "cr:Field",
720
+ "@id": "per-layer-kernel-summary/model",
721
+ "name": "model",
722
+ "dataType": "sc:Text",
723
+ "source": {
724
+ "fileObject": {
725
+ "@id": "per-layer-kernel-summary-parquet"
726
+ },
727
+ "extract": {
728
+ "column": "model"
729
+ }
730
+ }
731
+ },
732
+ {
733
+ "@type": "cr:Field",
734
+ "@id": "per-layer-kernel-summary/hardware",
735
+ "name": "hardware",
736
+ "dataType": "sc:Text",
737
+ "source": {
738
+ "fileObject": {
739
+ "@id": "per-layer-kernel-summary-parquet"
740
+ },
741
+ "extract": {
742
+ "column": "hardware"
743
+ }
744
+ }
745
+ },
746
+ {
747
+ "@type": "cr:Field",
748
+ "@id": "per-layer-kernel-summary/phase",
749
+ "name": "phase",
750
+ "dataType": "sc:Text",
751
+ "source": {
752
+ "fileObject": {
753
+ "@id": "per-layer-kernel-summary-parquet"
754
+ },
755
+ "extract": {
756
+ "column": "phase"
757
+ }
758
+ }
759
+ },
760
+ {
761
+ "@type": "cr:Field",
762
+ "@id": "per-layer-kernel-summary/batch_size",
763
+ "name": "batch_size",
764
+ "dataType": "sc:Integer",
765
+ "source": {
766
+ "fileObject": {
767
+ "@id": "per-layer-kernel-summary-parquet"
768
+ },
769
+ "extract": {
770
+ "column": "batch_size"
771
+ }
772
+ }
773
+ },
774
+ {
775
+ "@type": "cr:Field",
776
+ "@id": "per-layer-kernel-summary/sequence_length",
777
+ "name": "sequence_length",
778
+ "dataType": "sc:Integer",
779
+ "source": {
780
+ "fileObject": {
781
+ "@id": "per-layer-kernel-summary-parquet"
782
+ },
783
+ "extract": {
784
+ "column": "sequence_length"
785
+ }
786
+ }
787
+ },
788
+ {
789
+ "@type": "cr:Field",
790
+ "@id": "per-layer-kernel-summary/component_name",
791
+ "name": "component_name",
792
+ "dataType": "sc:Text",
793
+ "source": {
794
+ "fileObject": {
795
+ "@id": "per-layer-kernel-summary-parquet"
796
+ },
797
+ "extract": {
798
+ "column": "component_name"
799
+ }
800
+ }
801
+ },
802
+ {
803
+ "@type": "cr:Field",
804
+ "@id": "per-layer-kernel-summary/bound",
805
+ "name": "bound",
806
+ "dataType": "sc:Text",
807
+ "source": {
808
+ "fileObject": {
809
+ "@id": "per-layer-kernel-summary-parquet"
810
+ },
811
+ "extract": {
812
+ "column": "bound"
813
+ }
814
+ }
815
+ },
816
+ {
817
+ "@type": "cr:Field",
818
+ "@id": "per-layer-kernel-summary/flops",
819
+ "name": "flops",
820
+ "dataType": "sc:Float",
821
+ "source": {
822
+ "fileObject": {
823
+ "@id": "per-layer-kernel-summary-parquet"
824
+ },
825
+ "extract": {
826
+ "column": "flops"
827
+ }
828
+ }
829
+ },
830
+ {
831
+ "@type": "cr:Field",
832
+ "@id": "per-layer-kernel-summary/bytes_accessed",
833
+ "name": "bytes_accessed",
834
+ "dataType": "sc:Float",
835
+ "source": {
836
+ "fileObject": {
837
+ "@id": "per-layer-kernel-summary-parquet"
838
+ },
839
+ "extract": {
840
+ "column": "bytes_accessed"
841
+ }
842
+ }
843
+ },
844
+ {
845
+ "@type": "cr:Field",
846
+ "@id": "per-layer-kernel-summary/operational_intensity",
847
+ "name": "operational_intensity",
848
+ "dataType": "sc:Float",
849
+ "source": {
850
+ "fileObject": {
851
+ "@id": "per-layer-kernel-summary-parquet"
852
+ },
853
+ "extract": {
854
+ "column": "operational_intensity"
855
+ }
856
+ }
857
+ },
858
+ {
859
+ "@type": "cr:Field",
860
+ "@id": "per-layer-kernel-summary/kernel_name",
861
+ "name": "kernel_name",
862
+ "dataType": "sc:Text",
863
+ "source": {
864
+ "fileObject": {
865
+ "@id": "per-layer-kernel-summary-parquet"
866
+ },
867
+ "extract": {
868
+ "column": "kernel_name"
869
+ }
870
+ }
871
+ },
872
+ {
873
+ "@type": "cr:Field",
874
+ "@id": "per-layer-kernel-summary/duration_us",
875
+ "name": "duration_us",
876
+ "dataType": "sc:Float",
877
+ "source": {
878
+ "fileObject": {
879
+ "@id": "per-layer-kernel-summary-parquet"
880
+ },
881
+ "extract": {
882
+ "column": "duration_us"
883
+ }
884
+ }
885
+ },
886
+ {
887
+ "@type": "cr:Field",
888
+ "@id": "per-layer-kernel-summary/compute_sm_throughput_pct",
889
+ "name": "compute_sm_throughput_pct",
890
+ "dataType": "sc:Float",
891
+ "source": {
892
+ "fileObject": {
893
+ "@id": "per-layer-kernel-summary-parquet"
894
+ },
895
+ "extract": {
896
+ "column": "compute_sm_throughput_pct"
897
+ }
898
+ }
899
+ },
900
+ {
901
+ "@type": "cr:Field",
902
+ "@id": "per-layer-kernel-summary/dram_throughput_pct",
903
+ "name": "dram_throughput_pct",
904
+ "dataType": "sc:Float",
905
+ "source": {
906
+ "fileObject": {
907
+ "@id": "per-layer-kernel-summary-parquet"
908
+ },
909
+ "extract": {
910
+ "column": "dram_throughput_pct"
911
+ }
912
+ }
913
+ }
914
+ ]
915
+ }
916
+ ],
917
+ "rai:dataLimitations": "Results cover NVIDIA H100, A100, RTX 3090, and RTX 2080 Ti GPUs only and may not generalize to other accelerators (AMD, Intel, TPU). Benchmark configurations are pinned to vLLM 0.19.0 and SGLang 0.5.9; results do not represent other engine versions. Concurrency levels (1-320) may not cover extreme-scale deployments. Not recommended as sole basis for hardware purchasing decisions or for comparing model task quality.",
918
+ "rai:dataBiases": "Model selection over-represents Meta Llama and Alibaba Qwen families. Hardware is exclusively NVIDIA GPUs. Workload profiles are author-designed approximations of production traffic; real deployment patterns may differ.",
919
+ "rai:personalSensitiveInformation": "No personally identifiable information is present. All API endpoints and credentials are stripped. Workload traces use synthetic random tokens or publicly available coding benchmarks.",
920
+ "rai:dataUseCases": "Established uses: relative comparison of inference engine throughput, latency benchmarking under controlled conditions, TTFT scaling with context length in multi-turn sessions, per-kernel performance modelling. Not established: absolute latency prediction for production, model quality comparison, cost estimation.",
921
+ "rai:dataSocialImpact": "Enables reproducible comparison of open-source LLM serving systems, supporting infrastructure research and reducing vendor lock-in.",
922
+ "rai:hasSyntheticData": true,
923
+ "prov:wasDerivedFrom": [
924
+ {
925
+ "@id": "https://huggingface.co/datasets/princeton-nlp/SWE-bench",
926
+ "name": "SWE-Bench"
927
+ },
928
+ {
929
+ "@id": "https://github.com/TerminalBench/TerminalBench",
930
+ "name": "TerminalBench"
931
+ },
932
+ {
933
+ "@id": "https://github.com/xlang-ai/OSWorld",
934
+ "name": "OSWorld"
935
+ },
936
+ {
937
+ "@id": "https://huggingface.co/datasets/Aeala/ShareGPT_Vicuna_unfiltered",
938
+ "name": "ShareGPT"
939
+ }
940
+ ],
941
+ "prov:wasGeneratedBy": {
942
+ "@type": "prov:Activity",
943
+ "name": "AgentPerfBench benchmark collection",
944
+ "description": "Deploy model on target GPU with specified engine and tensor parallelism. Send requests per configuration after warmup using closed-loop concurrency control. Record per-request TTFT, TPOT, ITL, E2EL, and token counts. Compute summary statistics. Sanitize credentials and convert to Parquet."
945
+ },
946
+ "rai:dataCollection": "Automated benchmark collection. Models deployed on target GPU with specified engine and tensor parallelism. Requests sent per workload profile using closed-loop concurrency control with semaphore. Per-request TTFT, TPOT, ITL, E2EL, and token counts recorded. Kernel profiles collected via Nsight Compute (ncu) with hardware counter metrics.",
947
+ "rai:dataCollectionType": "Automated measurement via benchmarking scripts. No human subjects or crowdsourcing.",
948
+ "rai:dataCollectionMissingData": "Some model-hardware-concurrency combinations are absent due to OOM or engine incompatibility. Concurrency levels exceeding the session-pool size were filtered (see Concurrency filtering in README).",
949
+ "rai:dataPreprocessing": "Credentials and API endpoints stripped. Configurations with fewer than 75% of requests completing successfully are excluded. Summary statistics computed from successful requests. Kernel names demangled from NCU output.",
950
+ "rai:dataAnnotationProtocol": "Not applicable. Fully automated benchmark collection with no human annotation.",
951
+ "rai:dataAnnotationPlatform": "Not applicable. No human annotation.",
952
+ "rai:dataAnnotationAnalysis": "Not applicable. No human annotation."
953
+ }
kernel_profiles/kernels_labeled.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:3158723dcd2840b2c0080dbfc4d91e8004d7629743b679c9f970564099eb3649
3
- size 656666
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f9822c8c872bbfc00bc69d335ebef23bb8603f9a9e624267c56905f1e8cf8aa8
3
+ size 657407
synthetic_distributional/summary.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:0fd6ff9f802f9465a4618dfe8d5d426013028565e08ef5024f4302b9126a05a5
3
- size 64517
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7e641197686e02248ff063fceb9046efaf03e5e2e7fbb50d5d57f785e17c9157
3
+ size 307276