hf-transformers-bot commited on
Commit
f557761
·
verified ·
1 Parent(s): 6e02d9f

Upload benchmark results for run 18199493462

Browse files
2025-10-02/18199493462/benchmark_results/Llama-2-7b-hf/Llama-2-7b-hf_benchmark_20251002_165303.json ADDED
@@ -0,0 +1,1175 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model_name": "Llama-2-7b-hf",
3
+ "benchmark_scenarios": [
4
+ {
5
+ "scenario_name": "eager_eager_attn",
6
+ "metadata": {
7
+ "timestamp": "2025-10-02T16:47:01.786374",
8
+ "commit_id": "41e5abac5cb49983a08ddef3e8645d6efd23c8f3",
9
+ "hardware_info": {
10
+ "gpu_name": "NVIDIA A10G",
11
+ "gpu_memory_total_mb": 23028,
12
+ "cpu_count": 16,
13
+ "memory_total_mb": 63607,
14
+ "python_version": "3.10.12",
15
+ "torch_version": "2.8.0+cu126",
16
+ "cuda_version": "12.6"
17
+ },
18
+ "config": {
19
+ "name": "eager",
20
+ "model_id": "meta-llama/Llama-2-7b-hf",
21
+ "variant": "eager",
22
+ "warmup_iterations": 3,
23
+ "measurement_iterations": 5,
24
+ "num_tokens_to_generate": 100,
25
+ "device": "cuda",
26
+ "torch_dtype": "float16",
27
+ "compile_mode": null,
28
+ "compile_options": {},
29
+ "use_cache": true,
30
+ "batch_size": 1,
31
+ "sequence_length": null,
32
+ "attn_implementation": "eager",
33
+ "sdpa_backend": null,
34
+ "custom_params": {}
35
+ }
36
+ },
37
+ "measurements": {
38
+ "latency_seconds": {
39
+ "name": "latency_seconds",
40
+ "measurements": [
41
+ 3.6504755859375,
42
+ 3.65410986328125,
43
+ 3.651236572265625,
44
+ 3.65123828125,
45
+ 3.649072021484375
46
+ ],
47
+ "mean": 3.65122646484375,
48
+ "median": 3.651236572265625,
49
+ "std": 0.0016443645372859193,
50
+ "min": 3.649072021484375,
51
+ "max": 3.65410986328125,
52
+ "p25": 3.6504755859375,
53
+ "p75": 3.65123828125,
54
+ "p90": 3.65296123046875,
55
+ "p95": 3.653535546875,
56
+ "p99": 3.653995,
57
+ "unit": "seconds"
58
+ },
59
+ "time_to_first_token_seconds": {
60
+ "name": "time_to_first_token_seconds",
61
+ "measurements": [
62
+ 0.0376278076171875,
63
+ 0.03749001693725586,
64
+ 0.037065151214599606,
65
+ 0.03812787246704102,
66
+ 0.03705001449584961
67
+ ],
68
+ "mean": 0.03747217254638672,
69
+ "median": 0.03749001693725586,
70
+ "std": 0.0003995915083326358,
71
+ "min": 0.03705001449584961,
72
+ "max": 0.03812787246704102,
73
+ "p25": 0.037065151214599606,
74
+ "p75": 0.0376278076171875,
75
+ "p90": 0.03792784652709961,
76
+ "p95": 0.03802785949707031,
77
+ "p99": 0.03810786987304687,
78
+ "unit": "seconds"
79
+ },
80
+ "tokens_per_second": {
81
+ "name": "tokens_per_second",
82
+ "measurements": [
83
+ 27.39369094405775,
84
+ 27.36644593115869,
85
+ 27.387981583989532,
86
+ 27.3879687648775,
87
+ 27.404227543670636
88
+ ],
89
+ "mean": 27.388062953550822,
90
+ "median": 27.387981583989532,
91
+ "std": 0.012331094035009729,
92
+ "min": 27.36644593115869,
93
+ "max": 27.404227543670636,
94
+ "p25": 27.3879687648775,
95
+ "p75": 27.39369094405775,
96
+ "p90": 27.40001290382548,
97
+ "p95": 27.40212022374806,
98
+ "p99": 27.40380607968612,
99
+ "unit": "tokens/sec"
100
+ },
101
+ "time_per_output_token_seconds": {
102
+ "name": "time_per_output_token_seconds",
103
+ "measurements": [
104
+ 0.036504755859375,
105
+ 0.0365410986328125,
106
+ 0.03651236572265625,
107
+ 0.0365123828125,
108
+ 0.03649072021484375
109
+ ],
110
+ "mean": 0.0365122646484375,
111
+ "median": 0.03651236572265625,
112
+ "std": 1.6443645372857757e-05,
113
+ "min": 0.03649072021484375,
114
+ "max": 0.0365410986328125,
115
+ "p25": 0.036504755859375,
116
+ "p75": 0.0365123828125,
117
+ "p90": 0.0365296123046875,
118
+ "p95": 0.036535355468749996,
119
+ "p99": 0.036539949999999995,
120
+ "unit": "seconds/token"
121
+ }
122
+ },
123
+ "gpu_metrics": {
124
+ "gpu_utilization_mean": 91.06521739130434,
125
+ "gpu_utilization_max": 94,
126
+ "gpu_utilization_min": 88,
127
+ "gpu_memory_used_mean": 13253,
128
+ "gpu_memory_used_max": 13253,
129
+ "gpu_memory_used_min": 13253,
130
+ "sample_count": 92,
131
+ "gpu_monitoring_status": "success"
132
+ }
133
+ },
134
+ {
135
+ "scenario_name": "compiled_compile_max-autotune_eager_attn",
136
+ "metadata": {
137
+ "timestamp": "2025-10-02T16:47:37.088795",
138
+ "commit_id": "41e5abac5cb49983a08ddef3e8645d6efd23c8f3",
139
+ "hardware_info": {
140
+ "gpu_name": "NVIDIA A10G",
141
+ "gpu_memory_total_mb": 23028,
142
+ "cpu_count": 16,
143
+ "memory_total_mb": 63607,
144
+ "python_version": "3.10.12",
145
+ "torch_version": "2.8.0+cu126",
146
+ "cuda_version": "12.6"
147
+ },
148
+ "config": {
149
+ "name": "compiled",
150
+ "model_id": "meta-llama/Llama-2-7b-hf",
151
+ "variant": "compiled",
152
+ "warmup_iterations": 3,
153
+ "measurement_iterations": 5,
154
+ "num_tokens_to_generate": 100,
155
+ "device": "cuda",
156
+ "torch_dtype": "float16",
157
+ "compile_mode": "max-autotune",
158
+ "compile_options": {},
159
+ "use_cache": true,
160
+ "batch_size": 1,
161
+ "sequence_length": null,
162
+ "attn_implementation": "eager",
163
+ "sdpa_backend": null,
164
+ "custom_params": {}
165
+ }
166
+ },
167
+ "measurements": {
168
+ "latency_seconds": {
169
+ "name": "latency_seconds",
170
+ "measurements": [
171
+ 7.7810546875,
172
+ 7.80779931640625,
173
+ 7.85440478515625,
174
+ 7.859595703125,
175
+ 7.87478662109375
176
+ ],
177
+ "mean": 7.835528222656249,
178
+ "median": 7.85440478515625,
179
+ "std": 0.03525066907180073,
180
+ "min": 7.7810546875,
181
+ "max": 7.87478662109375,
182
+ "p25": 7.80779931640625,
183
+ "p75": 7.859595703125,
184
+ "p90": 7.86871025390625,
185
+ "p95": 7.8717484375,
186
+ "p99": 7.874178984375,
187
+ "unit": "seconds"
188
+ },
189
+ "time_to_first_token_seconds": {
190
+ "name": "time_to_first_token_seconds",
191
+ "measurements": [
192
+ 0.03877833557128906,
193
+ 0.03800835037231445,
194
+ 0.03837945556640625,
195
+ 0.03806140899658203,
196
+ 0.03803257751464844
197
+ ],
198
+ "mean": 0.03825202560424805,
199
+ "median": 0.03806140899658203,
200
+ "std": 0.00029567223422812726,
201
+ "min": 0.03800835037231445,
202
+ "max": 0.03877833557128906,
203
+ "p25": 0.03803257751464844,
204
+ "p75": 0.03837945556640625,
205
+ "p90": 0.038618783569335936,
206
+ "p95": 0.0386985595703125,
207
+ "p99": 0.03876238037109375,
208
+ "unit": "seconds"
209
+ },
210
+ "tokens_per_second": {
211
+ "name": "tokens_per_second",
212
+ "measurements": [
213
+ 12.851728206029268,
214
+ 12.807706236746323,
215
+ 12.731709497451202,
216
+ 12.723300762180386,
217
+ 12.698756780550168
218
+ ],
219
+ "mean": 12.76264029659147,
220
+ "median": 12.731709497451202,
221
+ "std": 0.05753783363140085,
222
+ "min": 12.698756780550168,
223
+ "max": 12.851728206029268,
224
+ "p25": 12.723300762180386,
225
+ "p75": 12.807706236746323,
226
+ "p90": 12.83411941831609,
227
+ "p95": 12.842923812172678,
228
+ "p99": 12.84996732725795,
229
+ "unit": "tokens/sec"
230
+ },
231
+ "time_per_output_token_seconds": {
232
+ "name": "time_per_output_token_seconds",
233
+ "measurements": [
234
+ 0.077810546875,
235
+ 0.0780779931640625,
236
+ 0.07854404785156249,
237
+ 0.07859595703125,
238
+ 0.0787478662109375
239
+ ],
240
+ "mean": 0.07835528222656249,
241
+ "median": 0.07854404785156249,
242
+ "std": 0.00035250669071800855,
243
+ "min": 0.077810546875,
244
+ "max": 0.0787478662109375,
245
+ "p25": 0.0780779931640625,
246
+ "p75": 0.07859595703125,
247
+ "p90": 0.0786871025390625,
248
+ "p95": 0.07871748437500001,
249
+ "p99": 0.07874178984375,
250
+ "unit": "seconds/token"
251
+ }
252
+ },
253
+ "gpu_metrics": {
254
+ "gpu_utilization_mean": 44.648809523809526,
255
+ "gpu_utilization_max": 94,
256
+ "gpu_utilization_min": 0,
257
+ "gpu_memory_used_mean": 13629.964285714286,
258
+ "gpu_memory_used_max": 13777,
259
+ "gpu_memory_used_min": 13415,
260
+ "sample_count": 168,
261
+ "gpu_monitoring_status": "success"
262
+ }
263
+ },
264
+ {
265
+ "scenario_name": "eager_sdpa_default",
266
+ "metadata": {
267
+ "timestamp": "2025-10-02T16:49:04.461647",
268
+ "commit_id": "41e5abac5cb49983a08ddef3e8645d6efd23c8f3",
269
+ "hardware_info": {
270
+ "gpu_name": "NVIDIA A10G",
271
+ "gpu_memory_total_mb": 23028,
272
+ "cpu_count": 16,
273
+ "memory_total_mb": 63607,
274
+ "python_version": "3.10.12",
275
+ "torch_version": "2.8.0+cu126",
276
+ "cuda_version": "12.6"
277
+ },
278
+ "config": {
279
+ "name": "eager",
280
+ "model_id": "meta-llama/Llama-2-7b-hf",
281
+ "variant": "eager",
282
+ "warmup_iterations": 3,
283
+ "measurement_iterations": 5,
284
+ "num_tokens_to_generate": 100,
285
+ "device": "cuda",
286
+ "torch_dtype": "float16",
287
+ "compile_mode": null,
288
+ "compile_options": {},
289
+ "use_cache": true,
290
+ "batch_size": 1,
291
+ "sequence_length": null,
292
+ "attn_implementation": "sdpa",
293
+ "sdpa_backend": null,
294
+ "custom_params": {}
295
+ }
296
+ },
297
+ "measurements": {
298
+ "latency_seconds": {
299
+ "name": "latency_seconds",
300
+ "measurements": [
301
+ 3.409737548828125,
302
+ 3.41694140625,
303
+ 3.4093505859375,
304
+ 3.409014892578125,
305
+ 3.410324951171875
306
+ ],
307
+ "mean": 3.4110738769531253,
308
+ "median": 3.409737548828125,
309
+ "std": 0.0029659302525244353,
310
+ "min": 3.409014892578125,
311
+ "max": 3.41694140625,
312
+ "p25": 3.4093505859375,
313
+ "p75": 3.410324951171875,
314
+ "p90": 3.41429482421875,
315
+ "p95": 3.4156181152343748,
316
+ "p99": 3.416676748046875,
317
+ "unit": "seconds"
318
+ },
319
+ "time_to_first_token_seconds": {
320
+ "name": "time_to_first_token_seconds",
321
+ "measurements": [
322
+ 0.03518080139160156,
323
+ 0.03468531036376953,
324
+ 0.03466134262084961,
325
+ 0.034588481903076174,
326
+ 0.03455110549926758
327
+ ],
328
+ "mean": 0.03473340835571289,
329
+ "median": 0.03466134262084961,
330
+ "std": 0.00022886914325615505,
331
+ "min": 0.03455110549926758,
332
+ "max": 0.03518080139160156,
333
+ "p25": 0.034588481903076174,
334
+ "p75": 0.03468531036376953,
335
+ "p90": 0.03498260498046875,
336
+ "p95": 0.03508170318603516,
337
+ "p99": 0.035160981750488284,
338
+ "unit": "seconds"
339
+ },
340
+ "tokens_per_second": {
341
+ "name": "tokens_per_second",
342
+ "measurements": [
343
+ 29.327770412819156,
344
+ 29.265939362345485,
345
+ 29.3310991285169,
346
+ 29.333987427779558,
347
+ 29.322718929067868
348
+ ],
349
+ "mean": 29.31630305210579,
350
+ "median": 29.327770412819156,
351
+ "std": 0.02545908916410651,
352
+ "min": 29.265939362345485,
353
+ "max": 29.333987427779558,
354
+ "p25": 29.322718929067868,
355
+ "p75": 29.3310991285169,
356
+ "p90": 29.332832108074495,
357
+ "p95": 29.333409767927026,
358
+ "p99": 29.33387189580905,
359
+ "unit": "tokens/sec"
360
+ },
361
+ "time_per_output_token_seconds": {
362
+ "name": "time_per_output_token_seconds",
363
+ "measurements": [
364
+ 0.03409737548828125,
365
+ 0.0341694140625,
366
+ 0.034093505859375005,
367
+ 0.03409014892578125,
368
+ 0.03410324951171875
369
+ ],
370
+ "mean": 0.03411073876953125,
371
+ "median": 0.03409737548828125,
372
+ "std": 2.9659302525244868e-05,
373
+ "min": 0.03409014892578125,
374
+ "max": 0.0341694140625,
375
+ "p25": 0.034093505859375005,
376
+ "p75": 0.03410324951171875,
377
+ "p90": 0.0341429482421875,
378
+ "p95": 0.03415618115234375,
379
+ "p99": 0.03416676748046875,
380
+ "unit": "seconds/token"
381
+ }
382
+ },
383
+ "gpu_metrics": {
384
+ "gpu_utilization_mean": 96.75581395348837,
385
+ "gpu_utilization_max": 97,
386
+ "gpu_utilization_min": 95,
387
+ "gpu_memory_used_mean": 13849,
388
+ "gpu_memory_used_max": 13849,
389
+ "gpu_memory_used_min": 13849,
390
+ "sample_count": 86,
391
+ "gpu_monitoring_status": "success"
392
+ }
393
+ },
394
+ {
395
+ "scenario_name": "eager_sdpa_math",
396
+ "metadata": {
397
+ "timestamp": "2025-10-02T16:49:37.034269",
398
+ "commit_id": "41e5abac5cb49983a08ddef3e8645d6efd23c8f3",
399
+ "hardware_info": {
400
+ "gpu_name": "NVIDIA A10G",
401
+ "gpu_memory_total_mb": 23028,
402
+ "cpu_count": 16,
403
+ "memory_total_mb": 63607,
404
+ "python_version": "3.10.12",
405
+ "torch_version": "2.8.0+cu126",
406
+ "cuda_version": "12.6"
407
+ },
408
+ "config": {
409
+ "name": "eager",
410
+ "model_id": "meta-llama/Llama-2-7b-hf",
411
+ "variant": "eager",
412
+ "warmup_iterations": 3,
413
+ "measurement_iterations": 5,
414
+ "num_tokens_to_generate": 100,
415
+ "device": "cuda",
416
+ "torch_dtype": "float16",
417
+ "compile_mode": null,
418
+ "compile_options": {},
419
+ "use_cache": true,
420
+ "batch_size": 1,
421
+ "sequence_length": null,
422
+ "attn_implementation": "sdpa",
423
+ "sdpa_backend": "math",
424
+ "custom_params": {}
425
+ }
426
+ },
427
+ "measurements": {
428
+ "latency_seconds": {
429
+ "name": "latency_seconds",
430
+ "measurements": [
431
+ 3.5586357421875,
432
+ 3.545120361328125,
433
+ 3.526301025390625,
434
+ 3.530726806640625,
435
+ 3.531763916015625
436
+ ],
437
+ "mean": 3.5385095703125,
438
+ "median": 3.531763916015625,
439
+ "std": 0.011864983637316699,
440
+ "min": 3.526301025390625,
441
+ "max": 3.5586357421875,
442
+ "p25": 3.530726806640625,
443
+ "p75": 3.545120361328125,
444
+ "p90": 3.5532295898437503,
445
+ "p95": 3.555932666015625,
446
+ "p99": 3.558095126953125,
447
+ "unit": "seconds"
448
+ },
449
+ "time_to_first_token_seconds": {
450
+ "name": "time_to_first_token_seconds",
451
+ "measurements": [
452
+ 0.04049814224243164,
453
+ 0.039877281188964844,
454
+ 0.03830886459350586,
455
+ 0.03775433731079102,
456
+ 0.03778188705444336
457
+ ],
458
+ "mean": 0.038844102478027345,
459
+ "median": 0.03830886459350586,
460
+ "std": 0.001131872352327191,
461
+ "min": 0.03775433731079102,
462
+ "max": 0.04049814224243164,
463
+ "p25": 0.03778188705444336,
464
+ "p75": 0.039877281188964844,
465
+ "p90": 0.04024979782104492,
466
+ "p95": 0.04037397003173828,
467
+ "p99": 0.040473307800292964,
468
+ "unit": "seconds"
469
+ },
470
+ "tokens_per_second": {
471
+ "name": "tokens_per_second",
472
+ "measurements": [
473
+ 28.10065633144285,
474
+ 28.207786988235437,
475
+ 28.3583276867075,
476
+ 28.322780400884895,
477
+ 28.3144633610775
478
+ ],
479
+ "mean": 28.260802953669632,
480
+ "median": 28.3144633610775,
481
+ "std": 0.0945403089080783,
482
+ "min": 28.10065633144285,
483
+ "max": 28.3583276867075,
484
+ "p25": 28.207786988235437,
485
+ "p75": 28.322780400884895,
486
+ "p90": 28.34410877237846,
487
+ "p95": 28.35121822954298,
488
+ "p99": 28.356905795274596,
489
+ "unit": "tokens/sec"
490
+ },
491
+ "time_per_output_token_seconds": {
492
+ "name": "time_per_output_token_seconds",
493
+ "measurements": [
494
+ 0.035586357421875,
495
+ 0.03545120361328125,
496
+ 0.03526301025390625,
497
+ 0.03530726806640625,
498
+ 0.03531763916015625
499
+ ],
500
+ "mean": 0.035385095703125,
501
+ "median": 0.03531763916015625,
502
+ "std": 0.00011864983637316678,
503
+ "min": 0.03526301025390625,
504
+ "max": 0.035586357421875,
505
+ "p25": 0.03530726806640625,
506
+ "p75": 0.03545120361328125,
507
+ "p90": 0.0355322958984375,
508
+ "p95": 0.03555932666015625,
509
+ "p99": 0.03558095126953125,
510
+ "unit": "seconds/token"
511
+ }
512
+ },
513
+ "gpu_metrics": {
514
+ "gpu_utilization_mean": 96.01123595505618,
515
+ "gpu_utilization_max": 98,
516
+ "gpu_utilization_min": 92,
517
+ "gpu_memory_used_mean": 13849,
518
+ "gpu_memory_used_max": 13849,
519
+ "gpu_memory_used_min": 13849,
520
+ "sample_count": 89,
521
+ "gpu_monitoring_status": "success"
522
+ }
523
+ },
524
+ {
525
+ "scenario_name": "eager_sdpa_flash_attention",
526
+ "metadata": {
527
+ "timestamp": "2025-10-02T16:50:10.726678",
528
+ "commit_id": "41e5abac5cb49983a08ddef3e8645d6efd23c8f3",
529
+ "hardware_info": {
530
+ "gpu_name": "NVIDIA A10G",
531
+ "gpu_memory_total_mb": 23028,
532
+ "cpu_count": 16,
533
+ "memory_total_mb": 63607,
534
+ "python_version": "3.10.12",
535
+ "torch_version": "2.8.0+cu126",
536
+ "cuda_version": "12.6"
537
+ },
538
+ "config": {
539
+ "name": "eager",
540
+ "model_id": "meta-llama/Llama-2-7b-hf",
541
+ "variant": "eager",
542
+ "warmup_iterations": 3,
543
+ "measurement_iterations": 5,
544
+ "num_tokens_to_generate": 100,
545
+ "device": "cuda",
546
+ "torch_dtype": "float16",
547
+ "compile_mode": null,
548
+ "compile_options": {},
549
+ "use_cache": true,
550
+ "batch_size": 1,
551
+ "sequence_length": null,
552
+ "attn_implementation": "sdpa",
553
+ "sdpa_backend": "flash_attention",
554
+ "custom_params": {}
555
+ }
556
+ },
557
+ "measurements": {
558
+ "latency_seconds": {
559
+ "name": "latency_seconds",
560
+ "measurements": [
561
+ 3.410571044921875,
562
+ 3.406748046875,
563
+ 3.40940771484375,
564
+ 3.409642578125,
565
+ 3.410457763671875
566
+ ],
567
+ "mean": 3.4093654296875,
568
+ "median": 3.409642578125,
569
+ "std": 0.0013839070564741992,
570
+ "min": 3.406748046875,
571
+ "max": 3.410571044921875,
572
+ "p25": 3.40940771484375,
573
+ "p75": 3.410457763671875,
574
+ "p90": 3.410525732421875,
575
+ "p95": 3.410548388671875,
576
+ "p99": 3.410566513671875,
577
+ "unit": "seconds"
578
+ },
579
+ "time_to_first_token_seconds": {
580
+ "name": "time_to_first_token_seconds",
581
+ "measurements": [
582
+ 0.035273696899414064,
583
+ 0.03467756652832031,
584
+ 0.03457913589477539,
585
+ 0.03469478225708008,
586
+ 0.03465350341796875
587
+ ],
588
+ "mean": 0.03477573699951172,
589
+ "median": 0.03467756652832031,
590
+ "std": 0.0002520905458311502,
591
+ "min": 0.03457913589477539,
592
+ "max": 0.035273696899414064,
593
+ "p25": 0.03465350341796875,
594
+ "p75": 0.03469478225708008,
595
+ "p90": 0.03504213104248047,
596
+ "p95": 0.035157913970947266,
597
+ "p99": 0.035250540313720706,
598
+ "unit": "seconds"
599
+ },
600
+ "tokens_per_second": {
601
+ "name": "tokens_per_second",
602
+ "measurements": [
603
+ 29.32060311392536,
604
+ 29.35350622472058,
605
+ 29.330607649130314,
606
+ 29.328587295795412,
607
+ 29.32157702265013
608
+ ],
609
+ "mean": 29.33097626124436,
610
+ "median": 29.328587295795412,
611
+ "std": 0.011911256495430468,
612
+ "min": 29.32060311392536,
613
+ "max": 29.35350622472058,
614
+ "p25": 29.32157702265013,
615
+ "p75": 29.330607649130314,
616
+ "p90": 29.344346794484473,
617
+ "p95": 29.348926509602528,
618
+ "p99": 29.35259028169697,
619
+ "unit": "tokens/sec"
620
+ },
621
+ "time_per_output_token_seconds": {
622
+ "name": "time_per_output_token_seconds",
623
+ "measurements": [
624
+ 0.03410571044921875,
625
+ 0.03406748046875,
626
+ 0.0340940771484375,
627
+ 0.03409642578125,
628
+ 0.03410457763671875
629
+ ],
630
+ "mean": 0.034093654296875005,
631
+ "median": 0.03409642578125,
632
+ "std": 1.3839070564742035e-05,
633
+ "min": 0.03406748046875,
634
+ "max": 0.03410571044921875,
635
+ "p25": 0.0340940771484375,
636
+ "p75": 0.03410457763671875,
637
+ "p90": 0.03410525732421875,
638
+ "p95": 0.03410548388671875,
639
+ "p99": 0.03410566513671875,
640
+ "unit": "seconds/token"
641
+ }
642
+ },
643
+ "gpu_metrics": {
644
+ "gpu_utilization_mean": 96.79069767441861,
645
+ "gpu_utilization_max": 97,
646
+ "gpu_utilization_min": 95,
647
+ "gpu_memory_used_mean": 13849,
648
+ "gpu_memory_used_max": 13849,
649
+ "gpu_memory_used_min": 13849,
650
+ "sample_count": 86,
651
+ "gpu_monitoring_status": "success"
652
+ }
653
+ },
654
+ {
655
+ "scenario_name": "eager_sdpa_efficient_attention",
656
+ "metadata": {
657
+ "timestamp": "2025-10-02T16:50:43.295286",
658
+ "commit_id": "41e5abac5cb49983a08ddef3e8645d6efd23c8f3",
659
+ "hardware_info": {
660
+ "gpu_name": "NVIDIA A10G",
661
+ "gpu_memory_total_mb": 23028,
662
+ "cpu_count": 16,
663
+ "memory_total_mb": 63607,
664
+ "python_version": "3.10.12",
665
+ "torch_version": "2.8.0+cu126",
666
+ "cuda_version": "12.6"
667
+ },
668
+ "config": {
669
+ "name": "eager",
670
+ "model_id": "meta-llama/Llama-2-7b-hf",
671
+ "variant": "eager",
672
+ "warmup_iterations": 3,
673
+ "measurement_iterations": 5,
674
+ "num_tokens_to_generate": 100,
675
+ "device": "cuda",
676
+ "torch_dtype": "float16",
677
+ "compile_mode": null,
678
+ "compile_options": {},
679
+ "use_cache": true,
680
+ "batch_size": 1,
681
+ "sequence_length": null,
682
+ "attn_implementation": "sdpa",
683
+ "sdpa_backend": "efficient_attention",
684
+ "custom_params": {}
685
+ }
686
+ },
687
+ "measurements": {
688
+ "latency_seconds": {
689
+ "name": "latency_seconds",
690
+ "measurements": [
691
+ 3.404732421875,
692
+ 3.4047578125,
693
+ 3.40570654296875,
694
+ 3.405033203125,
695
+ 3.40523583984375
696
+ ],
697
+ "mean": 3.4050931640625004,
698
+ "median": 3.405033203125,
699
+ "std": 0.00035854733900193294,
700
+ "min": 3.404732421875,
701
+ "max": 3.40570654296875,
702
+ "p25": 3.4047578125,
703
+ "p75": 3.40523583984375,
704
+ "p90": 3.40551826171875,
705
+ "p95": 3.40561240234375,
706
+ "p99": 3.40568771484375,
707
+ "unit": "seconds"
708
+ },
709
+ "time_to_first_token_seconds": {
710
+ "name": "time_to_first_token_seconds",
711
+ "measurements": [
712
+ 0.0351802864074707,
713
+ 0.03465439987182617,
714
+ 0.03466944122314453,
715
+ 0.034734622955322265,
716
+ 0.03465644836425781
717
+ ],
718
+ "mean": 0.034779039764404295,
719
+ "median": 0.03466944122314453,
720
+ "std": 0.00020275468722528677,
721
+ "min": 0.03465439987182617,
722
+ "max": 0.0351802864074707,
723
+ "p25": 0.03465644836425781,
724
+ "p75": 0.034734622955322265,
725
+ "p90": 0.035002021026611325,
726
+ "p95": 0.03509115371704102,
727
+ "p99": 0.035162459869384766,
728
+ "unit": "seconds"
729
+ },
730
+ "tokens_per_second": {
731
+ "name": "tokens_per_second",
732
+ "measurements": [
733
+ 29.37088370219989,
734
+ 29.370664671909022,
735
+ 29.362482861729518,
736
+ 29.368289245527503,
737
+ 29.366541615099564
738
+ ],
739
+ "mean": 29.3677724192931,
740
+ "median": 29.368289245527503,
741
+ "std": 0.003092138873279671,
742
+ "min": 29.362482861729518,
743
+ "max": 29.37088370219989,
744
+ "p25": 29.366541615099564,
745
+ "p75": 29.370664671909022,
746
+ "p90": 29.370796090083545,
747
+ "p95": 29.370839896141717,
748
+ "p99": 29.370874940988255,
749
+ "unit": "tokens/sec"
750
+ },
751
+ "time_per_output_token_seconds": {
752
+ "name": "time_per_output_token_seconds",
753
+ "measurements": [
754
+ 0.03404732421875,
755
+ 0.034047578125000005,
756
+ 0.0340570654296875,
757
+ 0.03405033203125,
758
+ 0.034052358398437496
759
+ ],
760
+ "mean": 0.034050931640625004,
761
+ "median": 0.03405033203125,
762
+ "std": 3.5854733900180275e-06,
763
+ "min": 0.03404732421875,
764
+ "max": 0.0340570654296875,
765
+ "p25": 0.034047578125000005,
766
+ "p75": 0.034052358398437496,
767
+ "p90": 0.0340551826171875,
768
+ "p95": 0.034056124023437495,
769
+ "p99": 0.0340568771484375,
770
+ "unit": "seconds/token"
771
+ }
772
+ },
773
+ "gpu_metrics": {
774
+ "gpu_utilization_mean": 97.01162790697674,
775
+ "gpu_utilization_max": 98,
776
+ "gpu_utilization_min": 95,
777
+ "gpu_memory_used_mean": 13849,
778
+ "gpu_memory_used_max": 13849,
779
+ "gpu_memory_used_min": 13849,
780
+ "sample_count": 86,
781
+ "gpu_monitoring_status": "success"
782
+ }
783
+ },
784
+ {
785
+ "scenario_name": "compiled_compile_max-autotune_sdpa_default",
786
+ "metadata": {
787
+ "timestamp": "2025-10-02T16:51:15.790032",
788
+ "commit_id": "41e5abac5cb49983a08ddef3e8645d6efd23c8f3",
789
+ "hardware_info": {
790
+ "gpu_name": "NVIDIA A10G",
791
+ "gpu_memory_total_mb": 23028,
792
+ "cpu_count": 16,
793
+ "memory_total_mb": 63607,
794
+ "python_version": "3.10.12",
795
+ "torch_version": "2.8.0+cu126",
796
+ "cuda_version": "12.6"
797
+ },
798
+ "config": {
799
+ "name": "compiled",
800
+ "model_id": "meta-llama/Llama-2-7b-hf",
801
+ "variant": "compiled",
802
+ "warmup_iterations": 3,
803
+ "measurement_iterations": 5,
804
+ "num_tokens_to_generate": 100,
805
+ "device": "cuda",
806
+ "torch_dtype": "float16",
807
+ "compile_mode": "max-autotune",
808
+ "compile_options": {},
809
+ "use_cache": true,
810
+ "batch_size": 1,
811
+ "sequence_length": null,
812
+ "attn_implementation": "sdpa",
813
+ "sdpa_backend": null,
814
+ "custom_params": {}
815
+ }
816
+ },
817
+ "measurements": {
818
+ "latency_seconds": {
819
+ "name": "latency_seconds",
820
+ "measurements": [
821
+ 3.6881279296875,
822
+ 3.6925576171875,
823
+ 3.688215576171875,
824
+ 3.691098388671875,
825
+ 3.688100341796875
826
+ ],
827
+ "mean": 3.689619970703125,
828
+ "median": 3.688215576171875,
829
+ "std": 0.0018613580505590886,
830
+ "min": 3.688100341796875,
831
+ "max": 3.6925576171875,
832
+ "p25": 3.6881279296875,
833
+ "p75": 3.691098388671875,
834
+ "p90": 3.69197392578125,
835
+ "p95": 3.692265771484375,
836
+ "p99": 3.6924992480468752,
837
+ "unit": "seconds"
838
+ },
839
+ "time_to_first_token_seconds": {
840
+ "name": "time_to_first_token_seconds",
841
+ "measurements": [
842
+ 0.03784572982788086,
843
+ 0.03743967819213867,
844
+ 0.03780137634277344,
845
+ 0.03732515335083008,
846
+ 0.037324512481689456
847
+ ],
848
+ "mean": 0.03754729003906251,
849
+ "median": 0.03743967819213867,
850
+ "std": 0.00022986126912107722,
851
+ "min": 0.037324512481689456,
852
+ "max": 0.03784572982788086,
853
+ "p25": 0.03732515335083008,
854
+ "p75": 0.03780137634277344,
855
+ "p90": 0.037827988433837896,
856
+ "p95": 0.03783685913085938,
857
+ "p99": 0.037843955688476566,
858
+ "unit": "seconds"
859
+ },
860
+ "tokens_per_second": {
861
+ "name": "tokens_per_second",
862
+ "measurements": [
863
+ 27.114026928147563,
864
+ 27.081500241062376,
865
+ 27.1133825924008,
866
+ 27.092206565640165,
867
+ 27.11422974768607
868
+ ],
869
+ "mean": 27.103069214987396,
870
+ "median": 27.1133825924008,
871
+ "std": 0.013669034410002234,
872
+ "min": 27.081500241062376,
873
+ "max": 27.11422974768607,
874
+ "p25": 27.092206565640165,
875
+ "p75": 27.114026928147563,
876
+ "p90": 27.11414861987067,
877
+ "p95": 27.114189183778368,
878
+ "p99": 27.11422163490453,
879
+ "unit": "tokens/sec"
880
+ },
881
+ "time_per_output_token_seconds": {
882
+ "name": "time_per_output_token_seconds",
883
+ "measurements": [
884
+ 0.036881279296875,
885
+ 0.036925576171875,
886
+ 0.03688215576171875,
887
+ 0.03691098388671875,
888
+ 0.03688100341796875
889
+ ],
890
+ "mean": 0.03689619970703125,
891
+ "median": 0.03688215576171875,
892
+ "std": 1.8613580505591275e-05,
893
+ "min": 0.03688100341796875,
894
+ "max": 0.036925576171875,
895
+ "p25": 0.036881279296875,
896
+ "p75": 0.03691098388671875,
897
+ "p90": 0.0369197392578125,
898
+ "p95": 0.03692265771484375,
899
+ "p99": 0.03692499248046875,
900
+ "unit": "seconds/token"
901
+ }
902
+ },
903
+ "gpu_metrics": {
904
+ "gpu_utilization_mean": 90.40860215053763,
905
+ "gpu_utilization_max": 93,
906
+ "gpu_utilization_min": 88,
907
+ "gpu_memory_used_mean": 13845,
908
+ "gpu_memory_used_max": 13845,
909
+ "gpu_memory_used_min": 13845,
910
+ "sample_count": 93,
911
+ "gpu_monitoring_status": "success"
912
+ }
913
+ },
914
+ {
915
+ "scenario_name": "compiled_compile_max-autotune_sdpa_math",
916
+ "metadata": {
917
+ "timestamp": "2025-10-02T16:51:50.460125",
918
+ "commit_id": "41e5abac5cb49983a08ddef3e8645d6efd23c8f3",
919
+ "hardware_info": {
920
+ "gpu_name": "NVIDIA A10G",
921
+ "gpu_memory_total_mb": 23028,
922
+ "cpu_count": 16,
923
+ "memory_total_mb": 63607,
924
+ "python_version": "3.10.12",
925
+ "torch_version": "2.8.0+cu126",
926
+ "cuda_version": "12.6"
927
+ },
928
+ "config": {
929
+ "name": "compiled",
930
+ "model_id": "meta-llama/Llama-2-7b-hf",
931
+ "variant": "compiled",
932
+ "warmup_iterations": 3,
933
+ "measurement_iterations": 5,
934
+ "num_tokens_to_generate": 100,
935
+ "device": "cuda",
936
+ "torch_dtype": "float16",
937
+ "compile_mode": "max-autotune",
938
+ "compile_options": {},
939
+ "use_cache": true,
940
+ "batch_size": 1,
941
+ "sequence_length": null,
942
+ "attn_implementation": "sdpa",
943
+ "sdpa_backend": "math",
944
+ "custom_params": {}
945
+ }
946
+ },
947
+ "measurements": {
948
+ "latency_seconds": {
949
+ "name": "latency_seconds",
950
+ "measurements": [
951
+ 3.86823681640625,
952
+ 3.874332763671875,
953
+ 3.8843779296875,
954
+ 3.864953125,
955
+ 3.914990966796875
956
+ ],
957
+ "mean": 3.8813783203125,
958
+ "median": 3.874332763671875,
959
+ "std": 0.01806064060012358,
960
+ "min": 3.864953125,
961
+ "max": 3.914990966796875,
962
+ "p25": 3.86823681640625,
963
+ "p75": 3.8843779296875,
964
+ "p90": 3.902745751953125,
965
+ "p95": 3.908868359375,
966
+ "p99": 3.9137664453125,
967
+ "unit": "seconds"
968
+ },
969
+ "time_to_first_token_seconds": {
970
+ "name": "time_to_first_token_seconds",
971
+ "measurements": [
972
+ 0.040766719818115235,
973
+ 0.04183427047729492,
974
+ 0.04199932861328125,
975
+ 0.04001408004760742,
976
+ 0.03998742294311523
977
+ ],
978
+ "mean": 0.04092036437988281,
979
+ "median": 0.040766719818115235,
980
+ "std": 0.0008619426707811444,
981
+ "min": 0.03998742294311523,
982
+ "max": 0.04199932861328125,
983
+ "p25": 0.04001408004760742,
984
+ "p75": 0.04183427047729492,
985
+ "p90": 0.041933305358886716,
986
+ "p95": 0.04196631698608398,
987
+ "p99": 0.04199272628784179,
988
+ "unit": "seconds"
989
+ },
990
+ "tokens_per_second": {
991
+ "name": "tokens_per_second",
992
+ "measurements": [
993
+ 25.851571335000134,
994
+ 25.8108959916044,
995
+ 25.744147920242415,
996
+ 25.873535012148434,
997
+ 25.542843099282276
998
+ ],
999
+ "mean": 25.764598671655534,
1000
+ "median": 25.8108959916044,
1001
+ "std": 0.11930513768957166,
1002
+ "min": 25.542843099282276,
1003
+ "max": 25.873535012148434,
1004
+ "p25": 25.744147920242415,
1005
+ "p75": 25.851571335000134,
1006
+ "p90": 25.864749541289115,
1007
+ "p95": 25.869142276718772,
1008
+ "p99": 25.8726564650625,
1009
+ "unit": "tokens/sec"
1010
+ },
1011
+ "time_per_output_token_seconds": {
1012
+ "name": "time_per_output_token_seconds",
1013
+ "measurements": [
1014
+ 0.0386823681640625,
1015
+ 0.03874332763671875,
1016
+ 0.038843779296875,
1017
+ 0.03864953125,
1018
+ 0.039149909667968746
1019
+ ],
1020
+ "mean": 0.038813783203125,
1021
+ "median": 0.03874332763671875,
1022
+ "std": 0.0001806064060012345,
1023
+ "min": 0.03864953125,
1024
+ "max": 0.039149909667968746,
1025
+ "p25": 0.0386823681640625,
1026
+ "p75": 0.038843779296875,
1027
+ "p90": 0.039027457519531246,
1028
+ "p95": 0.03908868359375,
1029
+ "p99": 0.039137664453125,
1030
+ "unit": "seconds/token"
1031
+ }
1032
+ },
1033
+ "gpu_metrics": {
1034
+ "gpu_utilization_mean": 89.06122448979592,
1035
+ "gpu_utilization_max": 93,
1036
+ "gpu_utilization_min": 81,
1037
+ "gpu_memory_used_mean": 13845,
1038
+ "gpu_memory_used_max": 13845,
1039
+ "gpu_memory_used_min": 13845,
1040
+ "sample_count": 98,
1041
+ "gpu_monitoring_status": "success"
1042
+ }
1043
+ },
1044
+ {
1045
+ "scenario_name": "compiled_compile_max-autotune_sdpa_efficient_attention",
1046
+ "metadata": {
1047
+ "timestamp": "2025-10-02T16:52:31.366623",
1048
+ "commit_id": "41e5abac5cb49983a08ddef3e8645d6efd23c8f3",
1049
+ "hardware_info": {
1050
+ "gpu_name": "NVIDIA A10G",
1051
+ "gpu_memory_total_mb": 23028,
1052
+ "cpu_count": 16,
1053
+ "memory_total_mb": 63607,
1054
+ "python_version": "3.10.12",
1055
+ "torch_version": "2.8.0+cu126",
1056
+ "cuda_version": "12.6"
1057
+ },
1058
+ "config": {
1059
+ "name": "compiled",
1060
+ "model_id": "meta-llama/Llama-2-7b-hf",
1061
+ "variant": "compiled",
1062
+ "warmup_iterations": 3,
1063
+ "measurement_iterations": 5,
1064
+ "num_tokens_to_generate": 100,
1065
+ "device": "cuda",
1066
+ "torch_dtype": "float16",
1067
+ "compile_mode": "max-autotune",
1068
+ "compile_options": {},
1069
+ "use_cache": true,
1070
+ "batch_size": 1,
1071
+ "sequence_length": null,
1072
+ "attn_implementation": "sdpa",
1073
+ "sdpa_backend": "efficient_attention",
1074
+ "custom_params": {}
1075
+ }
1076
+ },
1077
+ "measurements": {
1078
+ "latency_seconds": {
1079
+ "name": "latency_seconds",
1080
+ "measurements": [
1081
+ 3.693599365234375,
1082
+ 3.692274169921875,
1083
+ 3.688901611328125,
1084
+ 3.695408447265625,
1085
+ 3.692753662109375
1086
+ ],
1087
+ "mean": 3.6925874511718755,
1088
+ "median": 3.692753662109375,
1089
+ "std": 0.002130407266320661,
1090
+ "min": 3.688901611328125,
1091
+ "max": 3.695408447265625,
1092
+ "p25": 3.692274169921875,
1093
+ "p75": 3.693599365234375,
1094
+ "p90": 3.694684814453125,
1095
+ "p95": 3.695046630859375,
1096
+ "p99": 3.695336083984375,
1097
+ "unit": "seconds"
1098
+ },
1099
+ "time_to_first_token_seconds": {
1100
+ "name": "time_to_first_token_seconds",
1101
+ "measurements": [
1102
+ 0.03829974365234375,
1103
+ 0.037422046661376954,
1104
+ 0.037421409606933596,
1105
+ 0.038350017547607425,
1106
+ 0.03740185546875
1107
+ ],
1108
+ "mean": 0.03777901458740234,
1109
+ "median": 0.037422046661376954,
1110
+ "std": 0.0004460402588751944,
1111
+ "min": 0.03740185546875,
1112
+ "max": 0.038350017547607425,
1113
+ "p25": 0.037421409606933596,
1114
+ "p75": 0.03829974365234375,
1115
+ "p90": 0.038329907989501956,
1116
+ "p95": 0.03833996276855469,
1117
+ "p99": 0.03834800659179688,
1118
+ "unit": "seconds"
1119
+ },
1120
+ "tokens_per_second": {
1121
+ "name": "tokens_per_second",
1122
+ "measurements": [
1123
+ 27.073862136007424,
1124
+ 27.083579224593688,
1125
+ 27.108340242231815,
1126
+ 27.060608164706082,
1127
+ 27.080062508929444
1128
+ ],
1129
+ "mean": 27.08129045529369,
1130
+ "median": 27.080062508929444,
1131
+ "std": 0.015629295326079276,
1132
+ "min": 27.060608164706082,
1133
+ "max": 27.108340242231815,
1134
+ "p25": 27.073862136007424,
1135
+ "p75": 27.083579224593688,
1136
+ "p90": 27.098435835176563,
1137
+ "p95": 27.10338803870419,
1138
+ "p99": 27.10734980152629,
1139
+ "unit": "tokens/sec"
1140
+ },
1141
+ "time_per_output_token_seconds": {
1142
+ "name": "time_per_output_token_seconds",
1143
+ "measurements": [
1144
+ 0.03693599365234375,
1145
+ 0.03692274169921875,
1146
+ 0.03688901611328125,
1147
+ 0.03695408447265625,
1148
+ 0.03692753662109375
1149
+ ],
1150
+ "mean": 0.03692587451171875,
1151
+ "median": 0.03692753662109375,
1152
+ "std": 2.1304072663207484e-05,
1153
+ "min": 0.03688901611328125,
1154
+ "max": 0.03695408447265625,
1155
+ "p25": 0.03692274169921875,
1156
+ "p75": 0.03693599365234375,
1157
+ "p90": 0.03694684814453125,
1158
+ "p95": 0.03695046630859375,
1159
+ "p99": 0.036953360839843755,
1160
+ "unit": "seconds/token"
1161
+ }
1162
+ },
1163
+ "gpu_metrics": {
1164
+ "gpu_utilization_mean": 90.36559139784946,
1165
+ "gpu_utilization_max": 93,
1166
+ "gpu_utilization_min": 88,
1167
+ "gpu_memory_used_mean": 13845,
1168
+ "gpu_memory_used_max": 13845,
1169
+ "gpu_memory_used_min": 13845,
1170
+ "sample_count": 93,
1171
+ "gpu_monitoring_status": "success"
1172
+ }
1173
+ }
1174
+ ]
1175
+ }