e2e: speedups re-measured with flash attention off

#11
by awu12 - opened
runs/e2e-llama-3.1-8b/measurements/e2e_llama_faoff_20261105.json ADDED
The diff for this file is too large to render. See raw diff
 
runs/e2e-llama-3.1-8b/measurements/e2e_llama_faon_control_20261105.json ADDED
@@ -0,0 +1,1102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
3
+ "pp": 512,
4
+ "tg": 128,
5
+ "reps": 3,
6
+ "fa": "on",
7
+ "ub": null,
8
+ "ppl_chunks": null,
9
+ "hook": {
10
+ "agent": {
11
+ "expected": 7,
12
+ "hit": 7
13
+ }
14
+ },
15
+ "builds": {
16
+ "stock": {
17
+ "manifest": null,
18
+ "env": {}
19
+ },
20
+ "norepack": {
21
+ "manifest": null,
22
+ "env": {}
23
+ },
24
+ "agent": {
25
+ "manifest": "manifest.json",
26
+ "env": {
27
+ "ARMBENCH_OVERRIDE_STRICT": "1"
28
+ }
29
+ }
30
+ },
31
+ "summary": {
32
+ "stock/t16/pp": {
33
+ "median": 97.86225,
34
+ "max": 97.933314,
35
+ "min": 97.788732,
36
+ "n": 3
37
+ },
38
+ "stock/t16/tg": {
39
+ "median": 32.882476,
40
+ "max": 32.945278,
41
+ "min": 32.801316,
42
+ "n": 3
43
+ },
44
+ "norepack/t16/pp": {
45
+ "median": 52.137211,
46
+ "max": 52.142293,
47
+ "min": 52.133991,
48
+ "n": 3,
49
+ "speedup_vs_stock_median": 0.5327612128272138
50
+ },
51
+ "norepack/t16/tg": {
52
+ "median": 25.795281,
53
+ "max": 25.816133,
54
+ "min": 25.696947,
55
+ "n": 3,
56
+ "speedup_vs_stock_median": 0.7844689371931725
57
+ },
58
+ "agent/t16/pp": {
59
+ "median": 127.717702,
60
+ "max": 127.806925,
61
+ "min": 127.683074,
62
+ "n": 3,
63
+ "speedup_vs_stock_median": 1.305076288354294
64
+ },
65
+ "agent/t16/tg": {
66
+ "median": 36.407352,
67
+ "max": 36.454766,
68
+ "min": 36.363452,
69
+ "n": 3,
70
+ "speedup_vs_stock_median": 1.1071961855913772
71
+ }
72
+ },
73
+ "perplexity": {},
74
+ "samples": [
75
+ {
76
+ "build": "stock",
77
+ "threads": 16,
78
+ "rep": 0,
79
+ "kind": "pp",
80
+ "n": 512,
81
+ "tok_per_s": 97.933314,
82
+ "raw": {
83
+ "build_commit": "b29c606",
84
+ "build_number": 1,
85
+ "cpu_info": "CPU",
86
+ "gpu_info": "",
87
+ "backends": "CPU",
88
+ "model_filename": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
89
+ "model_type": "llama 8B Q4_K - Medium",
90
+ "model_size": 4912898304,
91
+ "model_n_params": 8030261312,
92
+ "n_batch": 2048,
93
+ "n_ubatch": 512,
94
+ "n_threads": 16,
95
+ "cpu_mask": "0x0",
96
+ "cpu_strict": false,
97
+ "poll": 50,
98
+ "type_k": "f16",
99
+ "type_v": "f16",
100
+ "n_gpu_layers": -1,
101
+ "n_cpu_moe": 0,
102
+ "split_mode": "layer",
103
+ "main_gpu": 0,
104
+ "no_kv_offload": false,
105
+ "flash_attn": 1,
106
+ "devices": "auto",
107
+ "tensor_split": "0.00",
108
+ "tensor_buft_overrides": "none",
109
+ "load_mode": "auto",
110
+ "lazy_mode": "auto",
111
+ "embeddings": false,
112
+ "no_op_offload": 0,
113
+ "no_host": false,
114
+ "fit_target": 0,
115
+ "fit_min_ctx": 0,
116
+ "n_prompt": 512,
117
+ "n_gen": 0,
118
+ "n_depth": 0,
119
+ "test_time": "2026-11-05T05:51:04Z",
120
+ "avg_ns": 5228047327,
121
+ "stddev_ns": 0,
122
+ "avg_ts": 97.933314,
123
+ "stddev_ts": 0.0,
124
+ "samples_ns": [
125
+ 5228047327
126
+ ],
127
+ "samples_ts": [
128
+ 97.9333
129
+ ]
130
+ }
131
+ },
132
+ {
133
+ "build": "stock",
134
+ "threads": 16,
135
+ "rep": 0,
136
+ "kind": "tg",
137
+ "n": 128,
138
+ "tok_per_s": 32.801316,
139
+ "raw": {
140
+ "build_commit": "b29c606",
141
+ "build_number": 1,
142
+ "cpu_info": "CPU",
143
+ "gpu_info": "",
144
+ "backends": "CPU",
145
+ "model_filename": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
146
+ "model_type": "llama 8B Q4_K - Medium",
147
+ "model_size": 4912898304,
148
+ "model_n_params": 8030261312,
149
+ "n_batch": 2048,
150
+ "n_ubatch": 512,
151
+ "n_threads": 16,
152
+ "cpu_mask": "0x0",
153
+ "cpu_strict": false,
154
+ "poll": 50,
155
+ "type_k": "f16",
156
+ "type_v": "f16",
157
+ "n_gpu_layers": -1,
158
+ "n_cpu_moe": 0,
159
+ "split_mode": "layer",
160
+ "main_gpu": 0,
161
+ "no_kv_offload": false,
162
+ "flash_attn": 1,
163
+ "devices": "auto",
164
+ "tensor_split": "0.00",
165
+ "tensor_buft_overrides": "none",
166
+ "load_mode": "auto",
167
+ "lazy_mode": "auto",
168
+ "embeddings": false,
169
+ "no_op_offload": 0,
170
+ "no_host": false,
171
+ "fit_target": 0,
172
+ "fit_min_ctx": 0,
173
+ "n_prompt": 0,
174
+ "n_gen": 128,
175
+ "n_depth": 0,
176
+ "test_time": "2026-11-05T05:51:14Z",
177
+ "avg_ns": 3902282445,
178
+ "stddev_ns": 0,
179
+ "avg_ts": 32.801316,
180
+ "stddev_ts": 0.0,
181
+ "samples_ns": [
182
+ 3902282445
183
+ ],
184
+ "samples_ts": [
185
+ 32.8013
186
+ ]
187
+ }
188
+ },
189
+ {
190
+ "build": "norepack",
191
+ "threads": 16,
192
+ "rep": 0,
193
+ "kind": "pp",
194
+ "n": 512,
195
+ "tok_per_s": 52.137211,
196
+ "raw": {
197
+ "build_commit": "b29c606",
198
+ "build_number": 1,
199
+ "cpu_info": "CPU",
200
+ "gpu_info": "",
201
+ "backends": "CPU",
202
+ "model_filename": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
203
+ "model_type": "llama 8B Q4_K - Medium",
204
+ "model_size": 4912898304,
205
+ "model_n_params": 8030261312,
206
+ "n_batch": 2048,
207
+ "n_ubatch": 512,
208
+ "n_threads": 16,
209
+ "cpu_mask": "0x0",
210
+ "cpu_strict": false,
211
+ "poll": 50,
212
+ "type_k": "f16",
213
+ "type_v": "f16",
214
+ "n_gpu_layers": -1,
215
+ "n_cpu_moe": 0,
216
+ "split_mode": "layer",
217
+ "main_gpu": 0,
218
+ "no_kv_offload": false,
219
+ "flash_attn": 1,
220
+ "devices": "auto",
221
+ "tensor_split": "0.00",
222
+ "tensor_buft_overrides": "none",
223
+ "load_mode": "auto",
224
+ "lazy_mode": "auto",
225
+ "embeddings": false,
226
+ "no_op_offload": 0,
227
+ "no_host": false,
228
+ "fit_target": 0,
229
+ "fit_min_ctx": 0,
230
+ "n_prompt": 512,
231
+ "n_gen": 0,
232
+ "n_depth": 0,
233
+ "test_time": "2026-11-05T05:51:19Z",
234
+ "avg_ns": 9820241460,
235
+ "stddev_ns": 0,
236
+ "avg_ts": 52.137211,
237
+ "stddev_ts": 0.0,
238
+ "samples_ns": [
239
+ 9820241460
240
+ ],
241
+ "samples_ts": [
242
+ 52.1372
243
+ ]
244
+ }
245
+ },
246
+ {
247
+ "build": "norepack",
248
+ "threads": 16,
249
+ "rep": 0,
250
+ "kind": "tg",
251
+ "n": 128,
252
+ "tok_per_s": 25.816133,
253
+ "raw": {
254
+ "build_commit": "b29c606",
255
+ "build_number": 1,
256
+ "cpu_info": "CPU",
257
+ "gpu_info": "",
258
+ "backends": "CPU",
259
+ "model_filename": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
260
+ "model_type": "llama 8B Q4_K - Medium",
261
+ "model_size": 4912898304,
262
+ "model_n_params": 8030261312,
263
+ "n_batch": 2048,
264
+ "n_ubatch": 512,
265
+ "n_threads": 16,
266
+ "cpu_mask": "0x0",
267
+ "cpu_strict": false,
268
+ "poll": 50,
269
+ "type_k": "f16",
270
+ "type_v": "f16",
271
+ "n_gpu_layers": -1,
272
+ "n_cpu_moe": 0,
273
+ "split_mode": "layer",
274
+ "main_gpu": 0,
275
+ "no_kv_offload": false,
276
+ "flash_attn": 1,
277
+ "devices": "auto",
278
+ "tensor_split": "0.00",
279
+ "tensor_buft_overrides": "none",
280
+ "load_mode": "auto",
281
+ "lazy_mode": "auto",
282
+ "embeddings": false,
283
+ "no_op_offload": 0,
284
+ "no_host": false,
285
+ "fit_target": 0,
286
+ "fit_min_ctx": 0,
287
+ "n_prompt": 0,
288
+ "n_gen": 128,
289
+ "n_depth": 0,
290
+ "test_time": "2026-11-05T05:51:39Z",
291
+ "avg_ns": 4958139847,
292
+ "stddev_ns": 0,
293
+ "avg_ts": 25.816133,
294
+ "stddev_ts": 0.0,
295
+ "samples_ns": [
296
+ 4958139847
297
+ ],
298
+ "samples_ts": [
299
+ 25.8161
300
+ ]
301
+ }
302
+ },
303
+ {
304
+ "build": "agent",
305
+ "threads": 16,
306
+ "rep": 0,
307
+ "kind": "pp",
308
+ "n": 512,
309
+ "tok_per_s": 127.717702,
310
+ "raw": {
311
+ "build_commit": "b29c606",
312
+ "build_number": 1,
313
+ "cpu_info": "CPU",
314
+ "gpu_info": "",
315
+ "backends": "CPU",
316
+ "model_filename": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
317
+ "model_type": "llama 8B Q4_K - Medium",
318
+ "model_size": 4912898304,
319
+ "model_n_params": 8030261312,
320
+ "n_batch": 2048,
321
+ "n_ubatch": 512,
322
+ "n_threads": 16,
323
+ "cpu_mask": "0x0",
324
+ "cpu_strict": false,
325
+ "poll": 50,
326
+ "type_k": "f16",
327
+ "type_v": "f16",
328
+ "n_gpu_layers": -1,
329
+ "n_cpu_moe": 0,
330
+ "split_mode": "layer",
331
+ "main_gpu": 0,
332
+ "no_kv_offload": false,
333
+ "flash_attn": 1,
334
+ "devices": "auto",
335
+ "tensor_split": "0.00",
336
+ "tensor_buft_overrides": "none",
337
+ "load_mode": "auto",
338
+ "lazy_mode": "auto",
339
+ "embeddings": false,
340
+ "no_op_offload": 0,
341
+ "no_host": false,
342
+ "fit_target": 0,
343
+ "fit_min_ctx": 0,
344
+ "n_prompt": 512,
345
+ "n_gen": 0,
346
+ "n_depth": 0,
347
+ "test_time": "2026-11-05T05:51:45Z",
348
+ "avg_ns": 4008841325,
349
+ "stddev_ns": 0,
350
+ "avg_ts": 127.717702,
351
+ "stddev_ts": 0.0,
352
+ "samples_ns": [
353
+ 4008841325
354
+ ],
355
+ "samples_ts": [
356
+ 127.718
357
+ ]
358
+ }
359
+ },
360
+ {
361
+ "build": "agent",
362
+ "threads": 16,
363
+ "rep": 0,
364
+ "kind": "tg",
365
+ "n": 128,
366
+ "tok_per_s": 36.454766,
367
+ "raw": {
368
+ "build_commit": "b29c606",
369
+ "build_number": 1,
370
+ "cpu_info": "CPU",
371
+ "gpu_info": "",
372
+ "backends": "CPU",
373
+ "model_filename": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
374
+ "model_type": "llama 8B Q4_K - Medium",
375
+ "model_size": 4912898304,
376
+ "model_n_params": 8030261312,
377
+ "n_batch": 2048,
378
+ "n_ubatch": 512,
379
+ "n_threads": 16,
380
+ "cpu_mask": "0x0",
381
+ "cpu_strict": false,
382
+ "poll": 50,
383
+ "type_k": "f16",
384
+ "type_v": "f16",
385
+ "n_gpu_layers": -1,
386
+ "n_cpu_moe": 0,
387
+ "split_mode": "layer",
388
+ "main_gpu": 0,
389
+ "no_kv_offload": false,
390
+ "flash_attn": 1,
391
+ "devices": "auto",
392
+ "tensor_split": "0.00",
393
+ "tensor_buft_overrides": "none",
394
+ "load_mode": "auto",
395
+ "lazy_mode": "auto",
396
+ "embeddings": false,
397
+ "no_op_offload": 0,
398
+ "no_host": false,
399
+ "fit_target": 0,
400
+ "fit_min_ctx": 0,
401
+ "n_prompt": 0,
402
+ "n_gen": 128,
403
+ "n_depth": 0,
404
+ "test_time": "2026-11-05T05:51:53Z",
405
+ "avg_ns": 3511200682,
406
+ "stddev_ns": 0,
407
+ "avg_ts": 36.454766,
408
+ "stddev_ts": 0.0,
409
+ "samples_ns": [
410
+ 3511200682
411
+ ],
412
+ "samples_ts": [
413
+ 36.4548
414
+ ]
415
+ }
416
+ },
417
+ {
418
+ "build": "stock",
419
+ "threads": 16,
420
+ "rep": 1,
421
+ "kind": "pp",
422
+ "n": 512,
423
+ "tok_per_s": 97.788732,
424
+ "raw": {
425
+ "build_commit": "b29c606",
426
+ "build_number": 1,
427
+ "cpu_info": "CPU",
428
+ "gpu_info": "",
429
+ "backends": "CPU",
430
+ "model_filename": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
431
+ "model_type": "llama 8B Q4_K - Medium",
432
+ "model_size": 4912898304,
433
+ "model_n_params": 8030261312,
434
+ "n_batch": 2048,
435
+ "n_ubatch": 512,
436
+ "n_threads": 16,
437
+ "cpu_mask": "0x0",
438
+ "cpu_strict": false,
439
+ "poll": 50,
440
+ "type_k": "f16",
441
+ "type_v": "f16",
442
+ "n_gpu_layers": -1,
443
+ "n_cpu_moe": 0,
444
+ "split_mode": "layer",
445
+ "main_gpu": 0,
446
+ "no_kv_offload": false,
447
+ "flash_attn": 1,
448
+ "devices": "auto",
449
+ "tensor_split": "0.00",
450
+ "tensor_buft_overrides": "none",
451
+ "load_mode": "auto",
452
+ "lazy_mode": "auto",
453
+ "embeddings": false,
454
+ "no_op_offload": 0,
455
+ "no_host": false,
456
+ "fit_target": 0,
457
+ "fit_min_ctx": 0,
458
+ "n_prompt": 512,
459
+ "n_gen": 0,
460
+ "n_depth": 0,
461
+ "test_time": "2026-11-05T05:51:59Z",
462
+ "avg_ns": 5235777044,
463
+ "stddev_ns": 0,
464
+ "avg_ts": 97.788732,
465
+ "stddev_ts": 0.0,
466
+ "samples_ns": [
467
+ 5235777044
468
+ ],
469
+ "samples_ts": [
470
+ 97.7887
471
+ ]
472
+ }
473
+ },
474
+ {
475
+ "build": "stock",
476
+ "threads": 16,
477
+ "rep": 1,
478
+ "kind": "tg",
479
+ "n": 128,
480
+ "tok_per_s": 32.945278,
481
+ "raw": {
482
+ "build_commit": "b29c606",
483
+ "build_number": 1,
484
+ "cpu_info": "CPU",
485
+ "gpu_info": "",
486
+ "backends": "CPU",
487
+ "model_filename": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
488
+ "model_type": "llama 8B Q4_K - Medium",
489
+ "model_size": 4912898304,
490
+ "model_n_params": 8030261312,
491
+ "n_batch": 2048,
492
+ "n_ubatch": 512,
493
+ "n_threads": 16,
494
+ "cpu_mask": "0x0",
495
+ "cpu_strict": false,
496
+ "poll": 50,
497
+ "type_k": "f16",
498
+ "type_v": "f16",
499
+ "n_gpu_layers": -1,
500
+ "n_cpu_moe": 0,
501
+ "split_mode": "layer",
502
+ "main_gpu": 0,
503
+ "no_kv_offload": false,
504
+ "flash_attn": 1,
505
+ "devices": "auto",
506
+ "tensor_split": "0.00",
507
+ "tensor_buft_overrides": "none",
508
+ "load_mode": "auto",
509
+ "lazy_mode": "auto",
510
+ "embeddings": false,
511
+ "no_op_offload": 0,
512
+ "no_host": false,
513
+ "fit_target": 0,
514
+ "fit_min_ctx": 0,
515
+ "n_prompt": 0,
516
+ "n_gen": 128,
517
+ "n_depth": 0,
518
+ "test_time": "2026-11-05T05:52:10Z",
519
+ "avg_ns": 3885230508,
520
+ "stddev_ns": 0,
521
+ "avg_ts": 32.945278,
522
+ "stddev_ts": 0.0,
523
+ "samples_ns": [
524
+ 3885230508
525
+ ],
526
+ "samples_ts": [
527
+ 32.9453
528
+ ]
529
+ }
530
+ },
531
+ {
532
+ "build": "norepack",
533
+ "threads": 16,
534
+ "rep": 1,
535
+ "kind": "pp",
536
+ "n": 512,
537
+ "tok_per_s": 52.133991,
538
+ "raw": {
539
+ "build_commit": "b29c606",
540
+ "build_number": 1,
541
+ "cpu_info": "CPU",
542
+ "gpu_info": "",
543
+ "backends": "CPU",
544
+ "model_filename": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
545
+ "model_type": "llama 8B Q4_K - Medium",
546
+ "model_size": 4912898304,
547
+ "model_n_params": 8030261312,
548
+ "n_batch": 2048,
549
+ "n_ubatch": 512,
550
+ "n_threads": 16,
551
+ "cpu_mask": "0x0",
552
+ "cpu_strict": false,
553
+ "poll": 50,
554
+ "type_k": "f16",
555
+ "type_v": "f16",
556
+ "n_gpu_layers": -1,
557
+ "n_cpu_moe": 0,
558
+ "split_mode": "layer",
559
+ "main_gpu": 0,
560
+ "no_kv_offload": false,
561
+ "flash_attn": 1,
562
+ "devices": "auto",
563
+ "tensor_split": "0.00",
564
+ "tensor_buft_overrides": "none",
565
+ "load_mode": "auto",
566
+ "lazy_mode": "auto",
567
+ "embeddings": false,
568
+ "no_op_offload": 0,
569
+ "no_host": false,
570
+ "fit_target": 0,
571
+ "fit_min_ctx": 0,
572
+ "n_prompt": 512,
573
+ "n_gen": 0,
574
+ "n_depth": 0,
575
+ "test_time": "2026-11-05T05:52:15Z",
576
+ "avg_ns": 9820847939,
577
+ "stddev_ns": 0,
578
+ "avg_ts": 52.133991,
579
+ "stddev_ts": 0.0,
580
+ "samples_ns": [
581
+ 9820847939
582
+ ],
583
+ "samples_ts": [
584
+ 52.134
585
+ ]
586
+ }
587
+ },
588
+ {
589
+ "build": "norepack",
590
+ "threads": 16,
591
+ "rep": 1,
592
+ "kind": "tg",
593
+ "n": 128,
594
+ "tok_per_s": 25.696947,
595
+ "raw": {
596
+ "build_commit": "b29c606",
597
+ "build_number": 1,
598
+ "cpu_info": "CPU",
599
+ "gpu_info": "",
600
+ "backends": "CPU",
601
+ "model_filename": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
602
+ "model_type": "llama 8B Q4_K - Medium",
603
+ "model_size": 4912898304,
604
+ "model_n_params": 8030261312,
605
+ "n_batch": 2048,
606
+ "n_ubatch": 512,
607
+ "n_threads": 16,
608
+ "cpu_mask": "0x0",
609
+ "cpu_strict": false,
610
+ "poll": 50,
611
+ "type_k": "f16",
612
+ "type_v": "f16",
613
+ "n_gpu_layers": -1,
614
+ "n_cpu_moe": 0,
615
+ "split_mode": "layer",
616
+ "main_gpu": 0,
617
+ "no_kv_offload": false,
618
+ "flash_attn": 1,
619
+ "devices": "auto",
620
+ "tensor_split": "0.00",
621
+ "tensor_buft_overrides": "none",
622
+ "load_mode": "auto",
623
+ "lazy_mode": "auto",
624
+ "embeddings": false,
625
+ "no_op_offload": 0,
626
+ "no_host": false,
627
+ "fit_target": 0,
628
+ "fit_min_ctx": 0,
629
+ "n_prompt": 0,
630
+ "n_gen": 128,
631
+ "n_depth": 0,
632
+ "test_time": "2026-11-05T05:52:34Z",
633
+ "avg_ns": 4981136397,
634
+ "stddev_ns": 0,
635
+ "avg_ts": 25.696947,
636
+ "stddev_ts": 0.0,
637
+ "samples_ns": [
638
+ 4981136397
639
+ ],
640
+ "samples_ts": [
641
+ 25.6969
642
+ ]
643
+ }
644
+ },
645
+ {
646
+ "build": "agent",
647
+ "threads": 16,
648
+ "rep": 1,
649
+ "kind": "pp",
650
+ "n": 512,
651
+ "tok_per_s": 127.806925,
652
+ "raw": {
653
+ "build_commit": "b29c606",
654
+ "build_number": 1,
655
+ "cpu_info": "CPU",
656
+ "gpu_info": "",
657
+ "backends": "CPU",
658
+ "model_filename": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
659
+ "model_type": "llama 8B Q4_K - Medium",
660
+ "model_size": 4912898304,
661
+ "model_n_params": 8030261312,
662
+ "n_batch": 2048,
663
+ "n_ubatch": 512,
664
+ "n_threads": 16,
665
+ "cpu_mask": "0x0",
666
+ "cpu_strict": false,
667
+ "poll": 50,
668
+ "type_k": "f16",
669
+ "type_v": "f16",
670
+ "n_gpu_layers": -1,
671
+ "n_cpu_moe": 0,
672
+ "split_mode": "layer",
673
+ "main_gpu": 0,
674
+ "no_kv_offload": false,
675
+ "flash_attn": 1,
676
+ "devices": "auto",
677
+ "tensor_split": "0.00",
678
+ "tensor_buft_overrides": "none",
679
+ "load_mode": "auto",
680
+ "lazy_mode": "auto",
681
+ "embeddings": false,
682
+ "no_op_offload": 0,
683
+ "no_host": false,
684
+ "fit_target": 0,
685
+ "fit_min_ctx": 0,
686
+ "n_prompt": 512,
687
+ "n_gen": 0,
688
+ "n_depth": 0,
689
+ "test_time": "2026-11-05T05:52:40Z",
690
+ "avg_ns": 4006042695,
691
+ "stddev_ns": 0,
692
+ "avg_ts": 127.806925,
693
+ "stddev_ts": 0.0,
694
+ "samples_ns": [
695
+ 4006042695
696
+ ],
697
+ "samples_ts": [
698
+ 127.807
699
+ ]
700
+ }
701
+ },
702
+ {
703
+ "build": "agent",
704
+ "threads": 16,
705
+ "rep": 1,
706
+ "kind": "tg",
707
+ "n": 128,
708
+ "tok_per_s": 36.407352,
709
+ "raw": {
710
+ "build_commit": "b29c606",
711
+ "build_number": 1,
712
+ "cpu_info": "CPU",
713
+ "gpu_info": "",
714
+ "backends": "CPU",
715
+ "model_filename": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
716
+ "model_type": "llama 8B Q4_K - Medium",
717
+ "model_size": 4912898304,
718
+ "model_n_params": 8030261312,
719
+ "n_batch": 2048,
720
+ "n_ubatch": 512,
721
+ "n_threads": 16,
722
+ "cpu_mask": "0x0",
723
+ "cpu_strict": false,
724
+ "poll": 50,
725
+ "type_k": "f16",
726
+ "type_v": "f16",
727
+ "n_gpu_layers": -1,
728
+ "n_cpu_moe": 0,
729
+ "split_mode": "layer",
730
+ "main_gpu": 0,
731
+ "no_kv_offload": false,
732
+ "flash_attn": 1,
733
+ "devices": "auto",
734
+ "tensor_split": "0.00",
735
+ "tensor_buft_overrides": "none",
736
+ "load_mode": "auto",
737
+ "lazy_mode": "auto",
738
+ "embeddings": false,
739
+ "no_op_offload": 0,
740
+ "no_host": false,
741
+ "fit_target": 0,
742
+ "fit_min_ctx": 0,
743
+ "n_prompt": 0,
744
+ "n_gen": 128,
745
+ "n_depth": 0,
746
+ "test_time": "2026-11-05T05:52:48Z",
747
+ "avg_ns": 3515773378,
748
+ "stddev_ns": 0,
749
+ "avg_ts": 36.407352,
750
+ "stddev_ts": 0.0,
751
+ "samples_ns": [
752
+ 3515773378
753
+ ],
754
+ "samples_ts": [
755
+ 36.4074
756
+ ]
757
+ }
758
+ },
759
+ {
760
+ "build": "stock",
761
+ "threads": 16,
762
+ "rep": 2,
763
+ "kind": "pp",
764
+ "n": 512,
765
+ "tok_per_s": 97.86225,
766
+ "raw": {
767
+ "build_commit": "b29c606",
768
+ "build_number": 1,
769
+ "cpu_info": "CPU",
770
+ "gpu_info": "",
771
+ "backends": "CPU",
772
+ "model_filename": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
773
+ "model_type": "llama 8B Q4_K - Medium",
774
+ "model_size": 4912898304,
775
+ "model_n_params": 8030261312,
776
+ "n_batch": 2048,
777
+ "n_ubatch": 512,
778
+ "n_threads": 16,
779
+ "cpu_mask": "0x0",
780
+ "cpu_strict": false,
781
+ "poll": 50,
782
+ "type_k": "f16",
783
+ "type_v": "f16",
784
+ "n_gpu_layers": -1,
785
+ "n_cpu_moe": 0,
786
+ "split_mode": "layer",
787
+ "main_gpu": 0,
788
+ "no_kv_offload": false,
789
+ "flash_attn": 1,
790
+ "devices": "auto",
791
+ "tensor_split": "0.00",
792
+ "tensor_buft_overrides": "none",
793
+ "load_mode": "auto",
794
+ "lazy_mode": "auto",
795
+ "embeddings": false,
796
+ "no_op_offload": 0,
797
+ "no_host": false,
798
+ "fit_target": 0,
799
+ "fit_min_ctx": 0,
800
+ "n_prompt": 512,
801
+ "n_gen": 0,
802
+ "n_depth": 0,
803
+ "test_time": "2026-11-05T05:52:54Z",
804
+ "avg_ns": 5231843757,
805
+ "stddev_ns": 0,
806
+ "avg_ts": 97.86225,
807
+ "stddev_ts": 0.0,
808
+ "samples_ns": [
809
+ 5231843757
810
+ ],
811
+ "samples_ts": [
812
+ 97.8622
813
+ ]
814
+ }
815
+ },
816
+ {
817
+ "build": "stock",
818
+ "threads": 16,
819
+ "rep": 2,
820
+ "kind": "tg",
821
+ "n": 128,
822
+ "tok_per_s": 32.882476,
823
+ "raw": {
824
+ "build_commit": "b29c606",
825
+ "build_number": 1,
826
+ "cpu_info": "CPU",
827
+ "gpu_info": "",
828
+ "backends": "CPU",
829
+ "model_filename": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
830
+ "model_type": "llama 8B Q4_K - Medium",
831
+ "model_size": 4912898304,
832
+ "model_n_params": 8030261312,
833
+ "n_batch": 2048,
834
+ "n_ubatch": 512,
835
+ "n_threads": 16,
836
+ "cpu_mask": "0x0",
837
+ "cpu_strict": false,
838
+ "poll": 50,
839
+ "type_k": "f16",
840
+ "type_v": "f16",
841
+ "n_gpu_layers": -1,
842
+ "n_cpu_moe": 0,
843
+ "split_mode": "layer",
844
+ "main_gpu": 0,
845
+ "no_kv_offload": false,
846
+ "flash_attn": 1,
847
+ "devices": "auto",
848
+ "tensor_split": "0.00",
849
+ "tensor_buft_overrides": "none",
850
+ "load_mode": "auto",
851
+ "lazy_mode": "auto",
852
+ "embeddings": false,
853
+ "no_op_offload": 0,
854
+ "no_host": false,
855
+ "fit_target": 0,
856
+ "fit_min_ctx": 0,
857
+ "n_prompt": 0,
858
+ "n_gen": 128,
859
+ "n_depth": 0,
860
+ "test_time": "2026-11-05T05:53:05Z",
861
+ "avg_ns": 3892650951,
862
+ "stddev_ns": 0,
863
+ "avg_ts": 32.882476,
864
+ "stddev_ts": 0.0,
865
+ "samples_ns": [
866
+ 3892650951
867
+ ],
868
+ "samples_ts": [
869
+ 32.8825
870
+ ]
871
+ }
872
+ },
873
+ {
874
+ "build": "norepack",
875
+ "threads": 16,
876
+ "rep": 2,
877
+ "kind": "pp",
878
+ "n": 512,
879
+ "tok_per_s": 52.142293,
880
+ "raw": {
881
+ "build_commit": "b29c606",
882
+ "build_number": 1,
883
+ "cpu_info": "CPU",
884
+ "gpu_info": "",
885
+ "backends": "CPU",
886
+ "model_filename": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
887
+ "model_type": "llama 8B Q4_K - Medium",
888
+ "model_size": 4912898304,
889
+ "model_n_params": 8030261312,
890
+ "n_batch": 2048,
891
+ "n_ubatch": 512,
892
+ "n_threads": 16,
893
+ "cpu_mask": "0x0",
894
+ "cpu_strict": false,
895
+ "poll": 50,
896
+ "type_k": "f16",
897
+ "type_v": "f16",
898
+ "n_gpu_layers": -1,
899
+ "n_cpu_moe": 0,
900
+ "split_mode": "layer",
901
+ "main_gpu": 0,
902
+ "no_kv_offload": false,
903
+ "flash_attn": 1,
904
+ "devices": "auto",
905
+ "tensor_split": "0.00",
906
+ "tensor_buft_overrides": "none",
907
+ "load_mode": "auto",
908
+ "lazy_mode": "auto",
909
+ "embeddings": false,
910
+ "no_op_offload": 0,
911
+ "no_host": false,
912
+ "fit_target": 0,
913
+ "fit_min_ctx": 0,
914
+ "n_prompt": 512,
915
+ "n_gen": 0,
916
+ "n_depth": 0,
917
+ "test_time": "2026-11-05T05:53:10Z",
918
+ "avg_ns": 9819284249,
919
+ "stddev_ns": 0,
920
+ "avg_ts": 52.142293,
921
+ "stddev_ts": 0.0,
922
+ "samples_ns": [
923
+ 9819284249
924
+ ],
925
+ "samples_ts": [
926
+ 52.1423
927
+ ]
928
+ }
929
+ },
930
+ {
931
+ "build": "norepack",
932
+ "threads": 16,
933
+ "rep": 2,
934
+ "kind": "tg",
935
+ "n": 128,
936
+ "tok_per_s": 25.795281,
937
+ "raw": {
938
+ "build_commit": "b29c606",
939
+ "build_number": 1,
940
+ "cpu_info": "CPU",
941
+ "gpu_info": "",
942
+ "backends": "CPU",
943
+ "model_filename": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
944
+ "model_type": "llama 8B Q4_K - Medium",
945
+ "model_size": 4912898304,
946
+ "model_n_params": 8030261312,
947
+ "n_batch": 2048,
948
+ "n_ubatch": 512,
949
+ "n_threads": 16,
950
+ "cpu_mask": "0x0",
951
+ "cpu_strict": false,
952
+ "poll": 50,
953
+ "type_k": "f16",
954
+ "type_v": "f16",
955
+ "n_gpu_layers": -1,
956
+ "n_cpu_moe": 0,
957
+ "split_mode": "layer",
958
+ "main_gpu": 0,
959
+ "no_kv_offload": false,
960
+ "flash_attn": 1,
961
+ "devices": "auto",
962
+ "tensor_split": "0.00",
963
+ "tensor_buft_overrides": "none",
964
+ "load_mode": "auto",
965
+ "lazy_mode": "auto",
966
+ "embeddings": false,
967
+ "no_op_offload": 0,
968
+ "no_host": false,
969
+ "fit_target": 0,
970
+ "fit_min_ctx": 0,
971
+ "n_prompt": 0,
972
+ "n_gen": 128,
973
+ "n_depth": 0,
974
+ "test_time": "2026-11-05T05:53:29Z",
975
+ "avg_ns": 4962147924,
976
+ "stddev_ns": 0,
977
+ "avg_ts": 25.795281,
978
+ "stddev_ts": 0.0,
979
+ "samples_ns": [
980
+ 4962147924
981
+ ],
982
+ "samples_ts": [
983
+ 25.7953
984
+ ]
985
+ }
986
+ },
987
+ {
988
+ "build": "agent",
989
+ "threads": 16,
990
+ "rep": 2,
991
+ "kind": "pp",
992
+ "n": 512,
993
+ "tok_per_s": 127.683074,
994
+ "raw": {
995
+ "build_commit": "b29c606",
996
+ "build_number": 1,
997
+ "cpu_info": "CPU",
998
+ "gpu_info": "",
999
+ "backends": "CPU",
1000
+ "model_filename": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
1001
+ "model_type": "llama 8B Q4_K - Medium",
1002
+ "model_size": 4912898304,
1003
+ "model_n_params": 8030261312,
1004
+ "n_batch": 2048,
1005
+ "n_ubatch": 512,
1006
+ "n_threads": 16,
1007
+ "cpu_mask": "0x0",
1008
+ "cpu_strict": false,
1009
+ "poll": 50,
1010
+ "type_k": "f16",
1011
+ "type_v": "f16",
1012
+ "n_gpu_layers": -1,
1013
+ "n_cpu_moe": 0,
1014
+ "split_mode": "layer",
1015
+ "main_gpu": 0,
1016
+ "no_kv_offload": false,
1017
+ "flash_attn": 1,
1018
+ "devices": "auto",
1019
+ "tensor_split": "0.00",
1020
+ "tensor_buft_overrides": "none",
1021
+ "load_mode": "auto",
1022
+ "lazy_mode": "auto",
1023
+ "embeddings": false,
1024
+ "no_op_offload": 0,
1025
+ "no_host": false,
1026
+ "fit_target": 0,
1027
+ "fit_min_ctx": 0,
1028
+ "n_prompt": 512,
1029
+ "n_gen": 0,
1030
+ "n_depth": 0,
1031
+ "test_time": "2026-11-05T05:53:35Z",
1032
+ "avg_ns": 4009928509,
1033
+ "stddev_ns": 0,
1034
+ "avg_ts": 127.683074,
1035
+ "stddev_ts": 0.0,
1036
+ "samples_ns": [
1037
+ 4009928509
1038
+ ],
1039
+ "samples_ts": [
1040
+ 127.683
1041
+ ]
1042
+ }
1043
+ },
1044
+ {
1045
+ "build": "agent",
1046
+ "threads": 16,
1047
+ "rep": 2,
1048
+ "kind": "tg",
1049
+ "n": 128,
1050
+ "tok_per_s": 36.363452,
1051
+ "raw": {
1052
+ "build_commit": "b29c606",
1053
+ "build_number": 1,
1054
+ "cpu_info": "CPU",
1055
+ "gpu_info": "",
1056
+ "backends": "CPU",
1057
+ "model_filename": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
1058
+ "model_type": "llama 8B Q4_K - Medium",
1059
+ "model_size": 4912898304,
1060
+ "model_n_params": 8030261312,
1061
+ "n_batch": 2048,
1062
+ "n_ubatch": 512,
1063
+ "n_threads": 16,
1064
+ "cpu_mask": "0x0",
1065
+ "cpu_strict": false,
1066
+ "poll": 50,
1067
+ "type_k": "f16",
1068
+ "type_v": "f16",
1069
+ "n_gpu_layers": -1,
1070
+ "n_cpu_moe": 0,
1071
+ "split_mode": "layer",
1072
+ "main_gpu": 0,
1073
+ "no_kv_offload": false,
1074
+ "flash_attn": 1,
1075
+ "devices": "auto",
1076
+ "tensor_split": "0.00",
1077
+ "tensor_buft_overrides": "none",
1078
+ "load_mode": "auto",
1079
+ "lazy_mode": "auto",
1080
+ "embeddings": false,
1081
+ "no_op_offload": 0,
1082
+ "no_host": false,
1083
+ "fit_target": 0,
1084
+ "fit_min_ctx": 0,
1085
+ "n_prompt": 0,
1086
+ "n_gen": 128,
1087
+ "n_depth": 0,
1088
+ "test_time": "2026-11-05T05:53:43Z",
1089
+ "avg_ns": 3520017811,
1090
+ "stddev_ns": 0,
1091
+ "avg_ts": 36.363452,
1092
+ "stddev_ts": 0.0,
1093
+ "samples_ns": [
1094
+ 3520017811
1095
+ ],
1096
+ "samples_ts": [
1097
+ 36.3635
1098
+ ]
1099
+ }
1100
+ }
1101
+ ]
1102
+ }
runs/e2e-qwen3.5-4b/measurements/e2e_qwen_faoff_20261105.json ADDED
The diff for this file is too large to render. See raw diff
 
runs/e2e-qwen3.5-4b/measurements/e2e_qwen_faon_control_20261105.json ADDED
@@ -0,0 +1,1102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "Qwen3.5-4B-Q4_K_M.gguf",
3
+ "pp": 512,
4
+ "tg": 128,
5
+ "reps": 3,
6
+ "fa": "on",
7
+ "ub": null,
8
+ "ppl_chunks": null,
9
+ "hook": {
10
+ "agent": {
11
+ "expected": 11,
12
+ "hit": 11
13
+ }
14
+ },
15
+ "builds": {
16
+ "stock": {
17
+ "manifest": null,
18
+ "env": {}
19
+ },
20
+ "norepack": {
21
+ "manifest": null,
22
+ "env": {}
23
+ },
24
+ "agent": {
25
+ "manifest": "manifest_gatev2b.json",
26
+ "env": {
27
+ "ARMBENCH_OVERRIDE_STRICT": "1"
28
+ }
29
+ }
30
+ },
31
+ "summary": {
32
+ "stock/t16/pp": {
33
+ "median": 157.078064,
34
+ "max": 157.099371,
35
+ "min": 156.993051,
36
+ "n": 3
37
+ },
38
+ "stock/t16/tg": {
39
+ "median": 37.517177,
40
+ "max": 37.87215,
41
+ "min": 37.458328,
42
+ "n": 3
43
+ },
44
+ "norepack/t16/pp": {
45
+ "median": 85.581952,
46
+ "max": 85.591136,
47
+ "min": 85.575976,
48
+ "n": 3,
49
+ "speedup_vs_stock_median": 0.5448370690384877
50
+ },
51
+ "norepack/t16/tg": {
52
+ "median": 33.100112,
53
+ "max": 33.225114,
54
+ "min": 33.005136,
55
+ "n": 3,
56
+ "speedup_vs_stock_median": 0.8822655286670424
57
+ },
58
+ "agent/t16/pp": {
59
+ "median": 208.506233,
60
+ "max": 208.513249,
61
+ "min": 208.248237,
62
+ "n": 3,
63
+ "speedup_vs_stock_median": 1.3274051620600569
64
+ },
65
+ "agent/t16/tg": {
66
+ "median": 41.220247,
67
+ "max": 41.461205,
68
+ "min": 40.777398,
69
+ "n": 3,
70
+ "speedup_vs_stock_median": 1.0987033219477043
71
+ }
72
+ },
73
+ "perplexity": {},
74
+ "samples": [
75
+ {
76
+ "build": "stock",
77
+ "threads": 16,
78
+ "rep": 0,
79
+ "kind": "pp",
80
+ "n": 512,
81
+ "tok_per_s": 156.993051,
82
+ "raw": {
83
+ "build_commit": "b29c606",
84
+ "build_number": 1,
85
+ "cpu_info": "CPU",
86
+ "gpu_info": "",
87
+ "backends": "CPU",
88
+ "model_filename": "Qwen3.5-4B-Q4_K_M.gguf",
89
+ "model_type": "qwen35 4B Q4_K - Medium",
90
+ "model_size": 2729969664,
91
+ "model_n_params": 4205751296,
92
+ "n_batch": 2048,
93
+ "n_ubatch": 512,
94
+ "n_threads": 16,
95
+ "cpu_mask": "0x0",
96
+ "cpu_strict": false,
97
+ "poll": 50,
98
+ "type_k": "f16",
99
+ "type_v": "f16",
100
+ "n_gpu_layers": -1,
101
+ "n_cpu_moe": 0,
102
+ "split_mode": "layer",
103
+ "main_gpu": 0,
104
+ "no_kv_offload": false,
105
+ "flash_attn": 1,
106
+ "devices": "auto",
107
+ "tensor_split": "0.00",
108
+ "tensor_buft_overrides": "none",
109
+ "load_mode": "auto",
110
+ "lazy_mode": "auto",
111
+ "embeddings": false,
112
+ "no_op_offload": 0,
113
+ "no_host": false,
114
+ "fit_target": 0,
115
+ "fit_min_ctx": 0,
116
+ "n_prompt": 512,
117
+ "n_gen": 0,
118
+ "n_depth": 0,
119
+ "test_time": "2026-11-05T05:49:07Z",
120
+ "avg_ns": 3261290838,
121
+ "stddev_ns": 0,
122
+ "avg_ts": 156.993051,
123
+ "stddev_ts": 0.0,
124
+ "samples_ns": [
125
+ 3261290838
126
+ ],
127
+ "samples_ts": [
128
+ 156.993
129
+ ]
130
+ }
131
+ },
132
+ {
133
+ "build": "stock",
134
+ "threads": 16,
135
+ "rep": 0,
136
+ "kind": "tg",
137
+ "n": 128,
138
+ "tok_per_s": 37.517177,
139
+ "raw": {
140
+ "build_commit": "b29c606",
141
+ "build_number": 1,
142
+ "cpu_info": "CPU",
143
+ "gpu_info": "",
144
+ "backends": "CPU",
145
+ "model_filename": "Qwen3.5-4B-Q4_K_M.gguf",
146
+ "model_type": "qwen35 4B Q4_K - Medium",
147
+ "model_size": 2729969664,
148
+ "model_n_params": 4205751296,
149
+ "n_batch": 2048,
150
+ "n_ubatch": 512,
151
+ "n_threads": 16,
152
+ "cpu_mask": "0x0",
153
+ "cpu_strict": false,
154
+ "poll": 50,
155
+ "type_k": "f16",
156
+ "type_v": "f16",
157
+ "n_gpu_layers": -1,
158
+ "n_cpu_moe": 0,
159
+ "split_mode": "layer",
160
+ "main_gpu": 0,
161
+ "no_kv_offload": false,
162
+ "flash_attn": 1,
163
+ "devices": "auto",
164
+ "tensor_split": "0.00",
165
+ "tensor_buft_overrides": "none",
166
+ "load_mode": "auto",
167
+ "lazy_mode": "auto",
168
+ "embeddings": false,
169
+ "no_op_offload": 0,
170
+ "no_host": false,
171
+ "fit_target": 0,
172
+ "fit_min_ctx": 0,
173
+ "n_prompt": 0,
174
+ "n_gen": 128,
175
+ "n_depth": 0,
176
+ "test_time": "2026-11-05T05:49:14Z",
177
+ "avg_ns": 3411770541,
178
+ "stddev_ns": 0,
179
+ "avg_ts": 37.517177,
180
+ "stddev_ts": 0.0,
181
+ "samples_ns": [
182
+ 3411770541
183
+ ],
184
+ "samples_ts": [
185
+ 37.5172
186
+ ]
187
+ }
188
+ },
189
+ {
190
+ "build": "norepack",
191
+ "threads": 16,
192
+ "rep": 0,
193
+ "kind": "pp",
194
+ "n": 512,
195
+ "tok_per_s": 85.575976,
196
+ "raw": {
197
+ "build_commit": "b29c606",
198
+ "build_number": 1,
199
+ "cpu_info": "CPU",
200
+ "gpu_info": "",
201
+ "backends": "CPU",
202
+ "model_filename": "Qwen3.5-4B-Q4_K_M.gguf",
203
+ "model_type": "qwen35 4B Q4_K - Medium",
204
+ "model_size": 2729969664,
205
+ "model_n_params": 4205751296,
206
+ "n_batch": 2048,
207
+ "n_ubatch": 512,
208
+ "n_threads": 16,
209
+ "cpu_mask": "0x0",
210
+ "cpu_strict": false,
211
+ "poll": 50,
212
+ "type_k": "f16",
213
+ "type_v": "f16",
214
+ "n_gpu_layers": -1,
215
+ "n_cpu_moe": 0,
216
+ "split_mode": "layer",
217
+ "main_gpu": 0,
218
+ "no_kv_offload": false,
219
+ "flash_attn": 1,
220
+ "devices": "auto",
221
+ "tensor_split": "0.00",
222
+ "tensor_buft_overrides": "none",
223
+ "load_mode": "auto",
224
+ "lazy_mode": "auto",
225
+ "embeddings": false,
226
+ "no_op_offload": 0,
227
+ "no_host": false,
228
+ "fit_target": 0,
229
+ "fit_min_ctx": 0,
230
+ "n_prompt": 512,
231
+ "n_gen": 0,
232
+ "n_depth": 0,
233
+ "test_time": "2026-11-05T05:49:18Z",
234
+ "avg_ns": 5982987552,
235
+ "stddev_ns": 0,
236
+ "avg_ts": 85.575976,
237
+ "stddev_ts": 0.0,
238
+ "samples_ns": [
239
+ 5982987552
240
+ ],
241
+ "samples_ts": [
242
+ 85.576
243
+ ]
244
+ }
245
+ },
246
+ {
247
+ "build": "norepack",
248
+ "threads": 16,
249
+ "rep": 0,
250
+ "kind": "tg",
251
+ "n": 128,
252
+ "tok_per_s": 33.100112,
253
+ "raw": {
254
+ "build_commit": "b29c606",
255
+ "build_number": 1,
256
+ "cpu_info": "CPU",
257
+ "gpu_info": "",
258
+ "backends": "CPU",
259
+ "model_filename": "Qwen3.5-4B-Q4_K_M.gguf",
260
+ "model_type": "qwen35 4B Q4_K - Medium",
261
+ "model_size": 2729969664,
262
+ "model_n_params": 4205751296,
263
+ "n_batch": 2048,
264
+ "n_ubatch": 512,
265
+ "n_threads": 16,
266
+ "cpu_mask": "0x0",
267
+ "cpu_strict": false,
268
+ "poll": 50,
269
+ "type_k": "f16",
270
+ "type_v": "f16",
271
+ "n_gpu_layers": -1,
272
+ "n_cpu_moe": 0,
273
+ "split_mode": "layer",
274
+ "main_gpu": 0,
275
+ "no_kv_offload": false,
276
+ "flash_attn": 1,
277
+ "devices": "auto",
278
+ "tensor_split": "0.00",
279
+ "tensor_buft_overrides": "none",
280
+ "load_mode": "auto",
281
+ "lazy_mode": "auto",
282
+ "embeddings": false,
283
+ "no_op_offload": 0,
284
+ "no_host": false,
285
+ "fit_target": 0,
286
+ "fit_min_ctx": 0,
287
+ "n_prompt": 0,
288
+ "n_gen": 128,
289
+ "n_depth": 0,
290
+ "test_time": "2026-11-05T05:49:30Z",
291
+ "avg_ns": 3867056390,
292
+ "stddev_ns": 0,
293
+ "avg_ts": 33.100112,
294
+ "stddev_ts": 0.0,
295
+ "samples_ns": [
296
+ 3867056390
297
+ ],
298
+ "samples_ts": [
299
+ 33.1001
300
+ ]
301
+ }
302
+ },
303
+ {
304
+ "build": "agent",
305
+ "threads": 16,
306
+ "rep": 0,
307
+ "kind": "pp",
308
+ "n": 512,
309
+ "tok_per_s": 208.513249,
310
+ "raw": {
311
+ "build_commit": "b29c606",
312
+ "build_number": 1,
313
+ "cpu_info": "CPU",
314
+ "gpu_info": "",
315
+ "backends": "CPU",
316
+ "model_filename": "Qwen3.5-4B-Q4_K_M.gguf",
317
+ "model_type": "qwen35 4B Q4_K - Medium",
318
+ "model_size": 2729969664,
319
+ "model_n_params": 4205751296,
320
+ "n_batch": 2048,
321
+ "n_ubatch": 512,
322
+ "n_threads": 16,
323
+ "cpu_mask": "0x0",
324
+ "cpu_strict": false,
325
+ "poll": 50,
326
+ "type_k": "f16",
327
+ "type_v": "f16",
328
+ "n_gpu_layers": -1,
329
+ "n_cpu_moe": 0,
330
+ "split_mode": "layer",
331
+ "main_gpu": 0,
332
+ "no_kv_offload": false,
333
+ "flash_attn": 1,
334
+ "devices": "auto",
335
+ "tensor_split": "0.00",
336
+ "tensor_buft_overrides": "none",
337
+ "load_mode": "auto",
338
+ "lazy_mode": "auto",
339
+ "embeddings": false,
340
+ "no_op_offload": 0,
341
+ "no_host": false,
342
+ "fit_target": 0,
343
+ "fit_min_ctx": 0,
344
+ "n_prompt": 512,
345
+ "n_gen": 0,
346
+ "n_depth": 0,
347
+ "test_time": "2026-11-05T05:49:35Z",
348
+ "avg_ns": 2455479460,
349
+ "stddev_ns": 0,
350
+ "avg_ts": 208.513249,
351
+ "stddev_ts": 0.0,
352
+ "samples_ns": [
353
+ 2455479460
354
+ ],
355
+ "samples_ts": [
356
+ 208.513
357
+ ]
358
+ }
359
+ },
360
+ {
361
+ "build": "agent",
362
+ "threads": 16,
363
+ "rep": 0,
364
+ "kind": "tg",
365
+ "n": 128,
366
+ "tok_per_s": 41.461205,
367
+ "raw": {
368
+ "build_commit": "b29c606",
369
+ "build_number": 1,
370
+ "cpu_info": "CPU",
371
+ "gpu_info": "",
372
+ "backends": "CPU",
373
+ "model_filename": "Qwen3.5-4B-Q4_K_M.gguf",
374
+ "model_type": "qwen35 4B Q4_K - Medium",
375
+ "model_size": 2729969664,
376
+ "model_n_params": 4205751296,
377
+ "n_batch": 2048,
378
+ "n_ubatch": 512,
379
+ "n_threads": 16,
380
+ "cpu_mask": "0x0",
381
+ "cpu_strict": false,
382
+ "poll": 50,
383
+ "type_k": "f16",
384
+ "type_v": "f16",
385
+ "n_gpu_layers": -1,
386
+ "n_cpu_moe": 0,
387
+ "split_mode": "layer",
388
+ "main_gpu": 0,
389
+ "no_kv_offload": false,
390
+ "flash_attn": 1,
391
+ "devices": "auto",
392
+ "tensor_split": "0.00",
393
+ "tensor_buft_overrides": "none",
394
+ "load_mode": "auto",
395
+ "lazy_mode": "auto",
396
+ "embeddings": false,
397
+ "no_op_offload": 0,
398
+ "no_host": false,
399
+ "fit_target": 0,
400
+ "fit_min_ctx": 0,
401
+ "n_prompt": 0,
402
+ "n_gen": 128,
403
+ "n_depth": 0,
404
+ "test_time": "2026-11-05T05:49:40Z",
405
+ "avg_ns": 3087223347,
406
+ "stddev_ns": 0,
407
+ "avg_ts": 41.461205,
408
+ "stddev_ts": 0.0,
409
+ "samples_ns": [
410
+ 3087223347
411
+ ],
412
+ "samples_ts": [
413
+ 41.4612
414
+ ]
415
+ }
416
+ },
417
+ {
418
+ "build": "stock",
419
+ "threads": 16,
420
+ "rep": 1,
421
+ "kind": "pp",
422
+ "n": 512,
423
+ "tok_per_s": 157.078064,
424
+ "raw": {
425
+ "build_commit": "b29c606",
426
+ "build_number": 1,
427
+ "cpu_info": "CPU",
428
+ "gpu_info": "",
429
+ "backends": "CPU",
430
+ "model_filename": "Qwen3.5-4B-Q4_K_M.gguf",
431
+ "model_type": "qwen35 4B Q4_K - Medium",
432
+ "model_size": 2729969664,
433
+ "model_n_params": 4205751296,
434
+ "n_batch": 2048,
435
+ "n_ubatch": 512,
436
+ "n_threads": 16,
437
+ "cpu_mask": "0x0",
438
+ "cpu_strict": false,
439
+ "poll": 50,
440
+ "type_k": "f16",
441
+ "type_v": "f16",
442
+ "n_gpu_layers": -1,
443
+ "n_cpu_moe": 0,
444
+ "split_mode": "layer",
445
+ "main_gpu": 0,
446
+ "no_kv_offload": false,
447
+ "flash_attn": 1,
448
+ "devices": "auto",
449
+ "tensor_split": "0.00",
450
+ "tensor_buft_overrides": "none",
451
+ "load_mode": "auto",
452
+ "lazy_mode": "auto",
453
+ "embeddings": false,
454
+ "no_op_offload": 0,
455
+ "no_host": false,
456
+ "fit_target": 0,
457
+ "fit_min_ctx": 0,
458
+ "n_prompt": 512,
459
+ "n_gen": 0,
460
+ "n_depth": 0,
461
+ "test_time": "2026-11-05T05:49:45Z",
462
+ "avg_ns": 3259525796,
463
+ "stddev_ns": 0,
464
+ "avg_ts": 157.078064,
465
+ "stddev_ts": 0.0,
466
+ "samples_ns": [
467
+ 3259525796
468
+ ],
469
+ "samples_ts": [
470
+ 157.078
471
+ ]
472
+ }
473
+ },
474
+ {
475
+ "build": "stock",
476
+ "threads": 16,
477
+ "rep": 1,
478
+ "kind": "tg",
479
+ "n": 128,
480
+ "tok_per_s": 37.87215,
481
+ "raw": {
482
+ "build_commit": "b29c606",
483
+ "build_number": 1,
484
+ "cpu_info": "CPU",
485
+ "gpu_info": "",
486
+ "backends": "CPU",
487
+ "model_filename": "Qwen3.5-4B-Q4_K_M.gguf",
488
+ "model_type": "qwen35 4B Q4_K - Medium",
489
+ "model_size": 2729969664,
490
+ "model_n_params": 4205751296,
491
+ "n_batch": 2048,
492
+ "n_ubatch": 512,
493
+ "n_threads": 16,
494
+ "cpu_mask": "0x0",
495
+ "cpu_strict": false,
496
+ "poll": 50,
497
+ "type_k": "f16",
498
+ "type_v": "f16",
499
+ "n_gpu_layers": -1,
500
+ "n_cpu_moe": 0,
501
+ "split_mode": "layer",
502
+ "main_gpu": 0,
503
+ "no_kv_offload": false,
504
+ "flash_attn": 1,
505
+ "devices": "auto",
506
+ "tensor_split": "0.00",
507
+ "tensor_buft_overrides": "none",
508
+ "load_mode": "auto",
509
+ "lazy_mode": "auto",
510
+ "embeddings": false,
511
+ "no_op_offload": 0,
512
+ "no_host": false,
513
+ "fit_target": 0,
514
+ "fit_min_ctx": 0,
515
+ "n_prompt": 0,
516
+ "n_gen": 128,
517
+ "n_depth": 0,
518
+ "test_time": "2026-11-05T05:49:52Z",
519
+ "avg_ns": 3379792285,
520
+ "stddev_ns": 0,
521
+ "avg_ts": 37.87215,
522
+ "stddev_ts": 0.0,
523
+ "samples_ns": [
524
+ 3379792285
525
+ ],
526
+ "samples_ts": [
527
+ 37.8721
528
+ ]
529
+ }
530
+ },
531
+ {
532
+ "build": "norepack",
533
+ "threads": 16,
534
+ "rep": 1,
535
+ "kind": "pp",
536
+ "n": 512,
537
+ "tok_per_s": 85.581952,
538
+ "raw": {
539
+ "build_commit": "b29c606",
540
+ "build_number": 1,
541
+ "cpu_info": "CPU",
542
+ "gpu_info": "",
543
+ "backends": "CPU",
544
+ "model_filename": "Qwen3.5-4B-Q4_K_M.gguf",
545
+ "model_type": "qwen35 4B Q4_K - Medium",
546
+ "model_size": 2729969664,
547
+ "model_n_params": 4205751296,
548
+ "n_batch": 2048,
549
+ "n_ubatch": 512,
550
+ "n_threads": 16,
551
+ "cpu_mask": "0x0",
552
+ "cpu_strict": false,
553
+ "poll": 50,
554
+ "type_k": "f16",
555
+ "type_v": "f16",
556
+ "n_gpu_layers": -1,
557
+ "n_cpu_moe": 0,
558
+ "split_mode": "layer",
559
+ "main_gpu": 0,
560
+ "no_kv_offload": false,
561
+ "flash_attn": 1,
562
+ "devices": "auto",
563
+ "tensor_split": "0.00",
564
+ "tensor_buft_overrides": "none",
565
+ "load_mode": "auto",
566
+ "lazy_mode": "auto",
567
+ "embeddings": false,
568
+ "no_op_offload": 0,
569
+ "no_host": false,
570
+ "fit_target": 0,
571
+ "fit_min_ctx": 0,
572
+ "n_prompt": 512,
573
+ "n_gen": 0,
574
+ "n_depth": 0,
575
+ "test_time": "2026-11-05T05:49:56Z",
576
+ "avg_ns": 5982569798,
577
+ "stddev_ns": 0,
578
+ "avg_ts": 85.581952,
579
+ "stddev_ts": 0.0,
580
+ "samples_ns": [
581
+ 5982569798
582
+ ],
583
+ "samples_ts": [
584
+ 85.582
585
+ ]
586
+ }
587
+ },
588
+ {
589
+ "build": "norepack",
590
+ "threads": 16,
591
+ "rep": 1,
592
+ "kind": "tg",
593
+ "n": 128,
594
+ "tok_per_s": 33.005136,
595
+ "raw": {
596
+ "build_commit": "b29c606",
597
+ "build_number": 1,
598
+ "cpu_info": "CPU",
599
+ "gpu_info": "",
600
+ "backends": "CPU",
601
+ "model_filename": "Qwen3.5-4B-Q4_K_M.gguf",
602
+ "model_type": "qwen35 4B Q4_K - Medium",
603
+ "model_size": 2729969664,
604
+ "model_n_params": 4205751296,
605
+ "n_batch": 2048,
606
+ "n_ubatch": 512,
607
+ "n_threads": 16,
608
+ "cpu_mask": "0x0",
609
+ "cpu_strict": false,
610
+ "poll": 50,
611
+ "type_k": "f16",
612
+ "type_v": "f16",
613
+ "n_gpu_layers": -1,
614
+ "n_cpu_moe": 0,
615
+ "split_mode": "layer",
616
+ "main_gpu": 0,
617
+ "no_kv_offload": false,
618
+ "flash_attn": 1,
619
+ "devices": "auto",
620
+ "tensor_split": "0.00",
621
+ "tensor_buft_overrides": "none",
622
+ "load_mode": "auto",
623
+ "lazy_mode": "auto",
624
+ "embeddings": false,
625
+ "no_op_offload": 0,
626
+ "no_host": false,
627
+ "fit_target": 0,
628
+ "fit_min_ctx": 0,
629
+ "n_prompt": 0,
630
+ "n_gen": 128,
631
+ "n_depth": 0,
632
+ "test_time": "2026-11-05T05:50:08Z",
633
+ "avg_ns": 3878184315,
634
+ "stddev_ns": 0,
635
+ "avg_ts": 33.005136,
636
+ "stddev_ts": 0.0,
637
+ "samples_ns": [
638
+ 3878184315
639
+ ],
640
+ "samples_ts": [
641
+ 33.0051
642
+ ]
643
+ }
644
+ },
645
+ {
646
+ "build": "agent",
647
+ "threads": 16,
648
+ "rep": 1,
649
+ "kind": "pp",
650
+ "n": 512,
651
+ "tok_per_s": 208.248237,
652
+ "raw": {
653
+ "build_commit": "b29c606",
654
+ "build_number": 1,
655
+ "cpu_info": "CPU",
656
+ "gpu_info": "",
657
+ "backends": "CPU",
658
+ "model_filename": "Qwen3.5-4B-Q4_K_M.gguf",
659
+ "model_type": "qwen35 4B Q4_K - Medium",
660
+ "model_size": 2729969664,
661
+ "model_n_params": 4205751296,
662
+ "n_batch": 2048,
663
+ "n_ubatch": 512,
664
+ "n_threads": 16,
665
+ "cpu_mask": "0x0",
666
+ "cpu_strict": false,
667
+ "poll": 50,
668
+ "type_k": "f16",
669
+ "type_v": "f16",
670
+ "n_gpu_layers": -1,
671
+ "n_cpu_moe": 0,
672
+ "split_mode": "layer",
673
+ "main_gpu": 0,
674
+ "no_kv_offload": false,
675
+ "flash_attn": 1,
676
+ "devices": "auto",
677
+ "tensor_split": "0.00",
678
+ "tensor_buft_overrides": "none",
679
+ "load_mode": "auto",
680
+ "lazy_mode": "auto",
681
+ "embeddings": false,
682
+ "no_op_offload": 0,
683
+ "no_host": false,
684
+ "fit_target": 0,
685
+ "fit_min_ctx": 0,
686
+ "n_prompt": 512,
687
+ "n_gen": 0,
688
+ "n_depth": 0,
689
+ "test_time": "2026-11-05T05:50:13Z",
690
+ "avg_ns": 2458604247,
691
+ "stddev_ns": 0,
692
+ "avg_ts": 208.248237,
693
+ "stddev_ts": 0.0,
694
+ "samples_ns": [
695
+ 2458604247
696
+ ],
697
+ "samples_ts": [
698
+ 208.248
699
+ ]
700
+ }
701
+ },
702
+ {
703
+ "build": "agent",
704
+ "threads": 16,
705
+ "rep": 1,
706
+ "kind": "tg",
707
+ "n": 128,
708
+ "tok_per_s": 40.777398,
709
+ "raw": {
710
+ "build_commit": "b29c606",
711
+ "build_number": 1,
712
+ "cpu_info": "CPU",
713
+ "gpu_info": "",
714
+ "backends": "CPU",
715
+ "model_filename": "Qwen3.5-4B-Q4_K_M.gguf",
716
+ "model_type": "qwen35 4B Q4_K - Medium",
717
+ "model_size": 2729969664,
718
+ "model_n_params": 4205751296,
719
+ "n_batch": 2048,
720
+ "n_ubatch": 512,
721
+ "n_threads": 16,
722
+ "cpu_mask": "0x0",
723
+ "cpu_strict": false,
724
+ "poll": 50,
725
+ "type_k": "f16",
726
+ "type_v": "f16",
727
+ "n_gpu_layers": -1,
728
+ "n_cpu_moe": 0,
729
+ "split_mode": "layer",
730
+ "main_gpu": 0,
731
+ "no_kv_offload": false,
732
+ "flash_attn": 1,
733
+ "devices": "auto",
734
+ "tensor_split": "0.00",
735
+ "tensor_buft_overrides": "none",
736
+ "load_mode": "auto",
737
+ "lazy_mode": "auto",
738
+ "embeddings": false,
739
+ "no_op_offload": 0,
740
+ "no_host": false,
741
+ "fit_target": 0,
742
+ "fit_min_ctx": 0,
743
+ "n_prompt": 0,
744
+ "n_gen": 128,
745
+ "n_depth": 0,
746
+ "test_time": "2026-11-05T05:50:17Z",
747
+ "avg_ns": 3138993841,
748
+ "stddev_ns": 0,
749
+ "avg_ts": 40.777398,
750
+ "stddev_ts": 0.0,
751
+ "samples_ns": [
752
+ 3138993841
753
+ ],
754
+ "samples_ts": [
755
+ 40.7774
756
+ ]
757
+ }
758
+ },
759
+ {
760
+ "build": "stock",
761
+ "threads": 16,
762
+ "rep": 2,
763
+ "kind": "pp",
764
+ "n": 512,
765
+ "tok_per_s": 157.099371,
766
+ "raw": {
767
+ "build_commit": "b29c606",
768
+ "build_number": 1,
769
+ "cpu_info": "CPU",
770
+ "gpu_info": "",
771
+ "backends": "CPU",
772
+ "model_filename": "Qwen3.5-4B-Q4_K_M.gguf",
773
+ "model_type": "qwen35 4B Q4_K - Medium",
774
+ "model_size": 2729969664,
775
+ "model_n_params": 4205751296,
776
+ "n_batch": 2048,
777
+ "n_ubatch": 512,
778
+ "n_threads": 16,
779
+ "cpu_mask": "0x0",
780
+ "cpu_strict": false,
781
+ "poll": 50,
782
+ "type_k": "f16",
783
+ "type_v": "f16",
784
+ "n_gpu_layers": -1,
785
+ "n_cpu_moe": 0,
786
+ "split_mode": "layer",
787
+ "main_gpu": 0,
788
+ "no_kv_offload": false,
789
+ "flash_attn": 1,
790
+ "devices": "auto",
791
+ "tensor_split": "0.00",
792
+ "tensor_buft_overrides": "none",
793
+ "load_mode": "auto",
794
+ "lazy_mode": "auto",
795
+ "embeddings": false,
796
+ "no_op_offload": 0,
797
+ "no_host": false,
798
+ "fit_target": 0,
799
+ "fit_min_ctx": 0,
800
+ "n_prompt": 512,
801
+ "n_gen": 0,
802
+ "n_depth": 0,
803
+ "test_time": "2026-11-05T05:50:23Z",
804
+ "avg_ns": 3259083706,
805
+ "stddev_ns": 0,
806
+ "avg_ts": 157.099371,
807
+ "stddev_ts": 0.0,
808
+ "samples_ns": [
809
+ 3259083706
810
+ ],
811
+ "samples_ts": [
812
+ 157.099
813
+ ]
814
+ }
815
+ },
816
+ {
817
+ "build": "stock",
818
+ "threads": 16,
819
+ "rep": 2,
820
+ "kind": "tg",
821
+ "n": 128,
822
+ "tok_per_s": 37.458328,
823
+ "raw": {
824
+ "build_commit": "b29c606",
825
+ "build_number": 1,
826
+ "cpu_info": "CPU",
827
+ "gpu_info": "",
828
+ "backends": "CPU",
829
+ "model_filename": "Qwen3.5-4B-Q4_K_M.gguf",
830
+ "model_type": "qwen35 4B Q4_K - Medium",
831
+ "model_size": 2729969664,
832
+ "model_n_params": 4205751296,
833
+ "n_batch": 2048,
834
+ "n_ubatch": 512,
835
+ "n_threads": 16,
836
+ "cpu_mask": "0x0",
837
+ "cpu_strict": false,
838
+ "poll": 50,
839
+ "type_k": "f16",
840
+ "type_v": "f16",
841
+ "n_gpu_layers": -1,
842
+ "n_cpu_moe": 0,
843
+ "split_mode": "layer",
844
+ "main_gpu": 0,
845
+ "no_kv_offload": false,
846
+ "flash_attn": 1,
847
+ "devices": "auto",
848
+ "tensor_split": "0.00",
849
+ "tensor_buft_overrides": "none",
850
+ "load_mode": "auto",
851
+ "lazy_mode": "auto",
852
+ "embeddings": false,
853
+ "no_op_offload": 0,
854
+ "no_host": false,
855
+ "fit_target": 0,
856
+ "fit_min_ctx": 0,
857
+ "n_prompt": 0,
858
+ "n_gen": 128,
859
+ "n_depth": 0,
860
+ "test_time": "2026-11-05T05:50:29Z",
861
+ "avg_ns": 3417130671,
862
+ "stddev_ns": 0,
863
+ "avg_ts": 37.458328,
864
+ "stddev_ts": 0.0,
865
+ "samples_ns": [
866
+ 3417130671
867
+ ],
868
+ "samples_ts": [
869
+ 37.4583
870
+ ]
871
+ }
872
+ },
873
+ {
874
+ "build": "norepack",
875
+ "threads": 16,
876
+ "rep": 2,
877
+ "kind": "pp",
878
+ "n": 512,
879
+ "tok_per_s": 85.591136,
880
+ "raw": {
881
+ "build_commit": "b29c606",
882
+ "build_number": 1,
883
+ "cpu_info": "CPU",
884
+ "gpu_info": "",
885
+ "backends": "CPU",
886
+ "model_filename": "Qwen3.5-4B-Q4_K_M.gguf",
887
+ "model_type": "qwen35 4B Q4_K - Medium",
888
+ "model_size": 2729969664,
889
+ "model_n_params": 4205751296,
890
+ "n_batch": 2048,
891
+ "n_ubatch": 512,
892
+ "n_threads": 16,
893
+ "cpu_mask": "0x0",
894
+ "cpu_strict": false,
895
+ "poll": 50,
896
+ "type_k": "f16",
897
+ "type_v": "f16",
898
+ "n_gpu_layers": -1,
899
+ "n_cpu_moe": 0,
900
+ "split_mode": "layer",
901
+ "main_gpu": 0,
902
+ "no_kv_offload": false,
903
+ "flash_attn": 1,
904
+ "devices": "auto",
905
+ "tensor_split": "0.00",
906
+ "tensor_buft_overrides": "none",
907
+ "load_mode": "auto",
908
+ "lazy_mode": "auto",
909
+ "embeddings": false,
910
+ "no_op_offload": 0,
911
+ "no_host": false,
912
+ "fit_target": 0,
913
+ "fit_min_ctx": 0,
914
+ "n_prompt": 512,
915
+ "n_gen": 0,
916
+ "n_depth": 0,
917
+ "test_time": "2026-11-05T05:50:34Z",
918
+ "avg_ns": 5981927873,
919
+ "stddev_ns": 0,
920
+ "avg_ts": 85.591136,
921
+ "stddev_ts": 0.0,
922
+ "samples_ns": [
923
+ 5981927873
924
+ ],
925
+ "samples_ts": [
926
+ 85.5911
927
+ ]
928
+ }
929
+ },
930
+ {
931
+ "build": "norepack",
932
+ "threads": 16,
933
+ "rep": 2,
934
+ "kind": "tg",
935
+ "n": 128,
936
+ "tok_per_s": 33.225114,
937
+ "raw": {
938
+ "build_commit": "b29c606",
939
+ "build_number": 1,
940
+ "cpu_info": "CPU",
941
+ "gpu_info": "",
942
+ "backends": "CPU",
943
+ "model_filename": "Qwen3.5-4B-Q4_K_M.gguf",
944
+ "model_type": "qwen35 4B Q4_K - Medium",
945
+ "model_size": 2729969664,
946
+ "model_n_params": 4205751296,
947
+ "n_batch": 2048,
948
+ "n_ubatch": 512,
949
+ "n_threads": 16,
950
+ "cpu_mask": "0x0",
951
+ "cpu_strict": false,
952
+ "poll": 50,
953
+ "type_k": "f16",
954
+ "type_v": "f16",
955
+ "n_gpu_layers": -1,
956
+ "n_cpu_moe": 0,
957
+ "split_mode": "layer",
958
+ "main_gpu": 0,
959
+ "no_kv_offload": false,
960
+ "flash_attn": 1,
961
+ "devices": "auto",
962
+ "tensor_split": "0.00",
963
+ "tensor_buft_overrides": "none",
964
+ "load_mode": "auto",
965
+ "lazy_mode": "auto",
966
+ "embeddings": false,
967
+ "no_op_offload": 0,
968
+ "no_host": false,
969
+ "fit_target": 0,
970
+ "fit_min_ctx": 0,
971
+ "n_prompt": 0,
972
+ "n_gen": 128,
973
+ "n_depth": 0,
974
+ "test_time": "2026-11-05T05:50:46Z",
975
+ "avg_ns": 3852507521,
976
+ "stddev_ns": 0,
977
+ "avg_ts": 33.225114,
978
+ "stddev_ts": 0.0,
979
+ "samples_ns": [
980
+ 3852507521
981
+ ],
982
+ "samples_ts": [
983
+ 33.2251
984
+ ]
985
+ }
986
+ },
987
+ {
988
+ "build": "agent",
989
+ "threads": 16,
990
+ "rep": 2,
991
+ "kind": "pp",
992
+ "n": 512,
993
+ "tok_per_s": 208.506233,
994
+ "raw": {
995
+ "build_commit": "b29c606",
996
+ "build_number": 1,
997
+ "cpu_info": "CPU",
998
+ "gpu_info": "",
999
+ "backends": "CPU",
1000
+ "model_filename": "Qwen3.5-4B-Q4_K_M.gguf",
1001
+ "model_type": "qwen35 4B Q4_K - Medium",
1002
+ "model_size": 2729969664,
1003
+ "model_n_params": 4205751296,
1004
+ "n_batch": 2048,
1005
+ "n_ubatch": 512,
1006
+ "n_threads": 16,
1007
+ "cpu_mask": "0x0",
1008
+ "cpu_strict": false,
1009
+ "poll": 50,
1010
+ "type_k": "f16",
1011
+ "type_v": "f16",
1012
+ "n_gpu_layers": -1,
1013
+ "n_cpu_moe": 0,
1014
+ "split_mode": "layer",
1015
+ "main_gpu": 0,
1016
+ "no_kv_offload": false,
1017
+ "flash_attn": 1,
1018
+ "devices": "auto",
1019
+ "tensor_split": "0.00",
1020
+ "tensor_buft_overrides": "none",
1021
+ "load_mode": "auto",
1022
+ "lazy_mode": "auto",
1023
+ "embeddings": false,
1024
+ "no_op_offload": 0,
1025
+ "no_host": false,
1026
+ "fit_target": 0,
1027
+ "fit_min_ctx": 0,
1028
+ "n_prompt": 512,
1029
+ "n_gen": 0,
1030
+ "n_depth": 0,
1031
+ "test_time": "2026-11-05T05:50:50Z",
1032
+ "avg_ns": 2455562079,
1033
+ "stddev_ns": 0,
1034
+ "avg_ts": 208.506233,
1035
+ "stddev_ts": 0.0,
1036
+ "samples_ns": [
1037
+ 2455562079
1038
+ ],
1039
+ "samples_ts": [
1040
+ 208.506
1041
+ ]
1042
+ }
1043
+ },
1044
+ {
1045
+ "build": "agent",
1046
+ "threads": 16,
1047
+ "rep": 2,
1048
+ "kind": "tg",
1049
+ "n": 128,
1050
+ "tok_per_s": 41.220247,
1051
+ "raw": {
1052
+ "build_commit": "b29c606",
1053
+ "build_number": 1,
1054
+ "cpu_info": "CPU",
1055
+ "gpu_info": "",
1056
+ "backends": "CPU",
1057
+ "model_filename": "Qwen3.5-4B-Q4_K_M.gguf",
1058
+ "model_type": "qwen35 4B Q4_K - Medium",
1059
+ "model_size": 2729969664,
1060
+ "model_n_params": 4205751296,
1061
+ "n_batch": 2048,
1062
+ "n_ubatch": 512,
1063
+ "n_threads": 16,
1064
+ "cpu_mask": "0x0",
1065
+ "cpu_strict": false,
1066
+ "poll": 50,
1067
+ "type_k": "f16",
1068
+ "type_v": "f16",
1069
+ "n_gpu_layers": -1,
1070
+ "n_cpu_moe": 0,
1071
+ "split_mode": "layer",
1072
+ "main_gpu": 0,
1073
+ "no_kv_offload": false,
1074
+ "flash_attn": 1,
1075
+ "devices": "auto",
1076
+ "tensor_split": "0.00",
1077
+ "tensor_buft_overrides": "none",
1078
+ "load_mode": "auto",
1079
+ "lazy_mode": "auto",
1080
+ "embeddings": false,
1081
+ "no_op_offload": 0,
1082
+ "no_host": false,
1083
+ "fit_target": 0,
1084
+ "fit_min_ctx": 0,
1085
+ "n_prompt": 0,
1086
+ "n_gen": 128,
1087
+ "n_depth": 0,
1088
+ "test_time": "2026-11-05T05:50:55Z",
1089
+ "avg_ns": 3105270056,
1090
+ "stddev_ns": 0,
1091
+ "avg_ts": 41.220247,
1092
+ "stddev_ts": 0.0,
1093
+ "samples_ns": [
1094
+ 3105270056
1095
+ ],
1096
+ "samples_ts": [
1097
+ 41.2202
1098
+ ]
1099
+ }
1100
+ }
1101
+ ]
1102
+ }
runs/e2e-qwen3.5-4b/measurements/manifest_gatev2b.json ADDED
@@ -0,0 +1,139 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "isa": "sve2",
3
+ "march": "-march=armv9-a+sve2",
4
+ "kernels": [
5
+ {
6
+ "op": "mul_mat",
7
+ "type": "q4_K",
8
+ "K": 2560,
9
+ "N": 1024,
10
+ "so": "gemm_ggml_q4_K_n1024_k2560.so",
11
+ "symbol": "armbench_entry_gemm",
12
+ "abi": "entry_rows",
13
+ "definition": "gemm_ggml_q4_K_n1024_k2560",
14
+ "version": "v18.cpp",
15
+ "harness_speedup": 3.107135600986274
16
+ },
17
+ {
18
+ "op": "mul_mat",
19
+ "type": "q4_K",
20
+ "K": 4096,
21
+ "N": 2560,
22
+ "so": "gemm_ggml_q4_K_n2560_k4096.so",
23
+ "symbol": "armbench_entry_gemm",
24
+ "abi": "entry_rows",
25
+ "definition": "gemm_ggml_q4_K_n2560_k4096",
26
+ "version": "v22.cpp",
27
+ "harness_speedup": 2.9470688355957457
28
+ },
29
+ {
30
+ "op": "mul_mat",
31
+ "type": "q4_K",
32
+ "K": 9216,
33
+ "N": 2560,
34
+ "so": "gemm_ggml_q4_K_n2560_k9216.so",
35
+ "symbol": "armbench_entry_gemm",
36
+ "abi": "entry_rows",
37
+ "definition": "gemm_ggml_q4_K_n2560_k9216",
38
+ "version": "v18.cpp",
39
+ "harness_speedup": 2.8688158216539787
40
+ },
41
+ {
42
+ "op": "mul_mat",
43
+ "type": "q4_K",
44
+ "K": 2560,
45
+ "N": 4096,
46
+ "so": "gemm_ggml_q4_K_n4096_k2560.so",
47
+ "symbol": "armbench_entry_gemm",
48
+ "abi": "entry_rows",
49
+ "definition": "gemm_ggml_q4_K_n4096_k2560",
50
+ "version": "v12.cpp",
51
+ "harness_speedup": 3.025973381483579
52
+ },
53
+ {
54
+ "op": "mul_mat",
55
+ "type": "q4_K",
56
+ "K": 2560,
57
+ "N": 8192,
58
+ "so": "gemm_ggml_q4_K_n8192_k2560.so",
59
+ "symbol": "armbench_entry_gemm",
60
+ "abi": "entry_rows",
61
+ "definition": "gemm_ggml_q4_K_n8192_k2560",
62
+ "version": "v7.cpp",
63
+ "harness_speedup": 2.927434264579252
64
+ },
65
+ {
66
+ "op": "mul_mat",
67
+ "type": "q4_K",
68
+ "K": 2560,
69
+ "N": 9216,
70
+ "so": "gemm_ggml_q4_K_n9216_k2560.so",
71
+ "symbol": "armbench_entry_gemm",
72
+ "abi": "entry_rows",
73
+ "definition": "gemm_ggml_q4_K_n9216_k2560",
74
+ "version": "v3.cpp",
75
+ "harness_speedup": 2.966901769113881
76
+ },
77
+ {
78
+ "op": "mul_mat",
79
+ "type": "q5_K",
80
+ "K": 4096,
81
+ "N": 2560,
82
+ "so": "gemm_ggml_q5_K_n2560_k4096.so",
83
+ "symbol": "armbench_entry_gemm",
84
+ "abi": "entry_rows",
85
+ "definition": "gemm_ggml_q5_K_n2560_k4096",
86
+ "version": "v10.cpp",
87
+ "harness_speedup": 3.600647529426703
88
+ },
89
+ {
90
+ "op": "mul_mat",
91
+ "type": "q5_K",
92
+ "K": 2560,
93
+ "N": 8192,
94
+ "so": "gemm_ggml_q5_K_n8192_k2560.so",
95
+ "symbol": "armbench_entry_gemm",
96
+ "abi": "entry_rows",
97
+ "definition": "gemm_ggml_q5_K_n8192_k2560",
98
+ "version": "v18.cpp",
99
+ "harness_speedup": 3.573490701041923
100
+ },
101
+ {
102
+ "op": "mul_mat",
103
+ "type": "q6_K",
104
+ "K": 2560,
105
+ "N": 1024,
106
+ "so": "gemm_ggml_q6_K_n1024_k2560.so",
107
+ "symbol": "armbench_entry_gemm",
108
+ "abi": "entry_rows",
109
+ "definition": "gemm_ggml_q6_K_n1024_k2560",
110
+ "version": "v14.cpp",
111
+ "harness_speedup": 2.704438384827637
112
+ },
113
+ {
114
+ "op": "mul_mat",
115
+ "type": "q6_K",
116
+ "K": 2560,
117
+ "N": 248320,
118
+ "so": "gemm_ggml_q6_K_n248320_k2560.so",
119
+ "symbol": "armbench_entry_gemm",
120
+ "abi": "entry_rows",
121
+ "definition": "gemm_ggml_q6_K_n248320_k2560",
122
+ "version": "v18.cpp",
123
+ "harness_speedup": 2.2496635860168706
124
+ },
125
+ {
126
+ "op": "mul_mat",
127
+ "type": "q6_K",
128
+ "K": 9216,
129
+ "N": 2560,
130
+ "so": "gemm_ggml_q6_K_n2560_k9216.so",
131
+ "symbol": "armbench_entry_gemm",
132
+ "abi": "entry_rows",
133
+ "definition": "gemm_ggml_q6_K_n2560_k9216",
134
+ "version": "v17.cpp",
135
+ "harness_speedup": 2.7407442871994214
136
+ }
137
+ ],
138
+ "skipped": []
139
+ }