strata-headquotient-q25 / artifacts /cached_decode.json
nur-dev's picture
Add files using upload-large-folder tool
9210292 verified
Raw
History Blame Contribute Delete
14.6 kB
{
"created_at": "2026-07-21T02:07:16.839792+00:00",
"device": "cuda:6",
"graph_enabled": false,
"measurement": "inference-only KV cache; BF16 numerical equivalence is checked against full-sequence logits; prefill and autoregressive decode are reported separately",
"program": "STRATA-HEADQUOTIENT-Q25-CACHED-DECODE-PROFILE",
"rows": [
{
"batch_size": 1,
"dense": {
"batch_size": 1,
"decode_seconds": 0.23088740557432175,
"decode_steps": 32,
"decode_tokens_per_second": 138.59569308426106,
"first_cached_token_ms": 7.441435630122821,
"peak_allocated_bytes": 2767192064,
"persistent_kv_bytes": 204472320,
"prefill_seconds": 0.02532536784807841,
"prefill_tokens_per_second": 80867.53220271169,
"prefix_length": 2048
},
"prefix_length": 2048,
"q25": {
"batch_size": 1,
"decode_seconds": 0.21997203739980856,
"decode_steps": 32,
"decode_tokens_per_second": 145.47303547422547,
"first_cached_token_ms": 7.406104976932208,
"peak_allocated_bytes": 2728734208,
"persistent_kv_bytes": 178520064,
"prefill_seconds": 0.02850375821193059,
"prefill_tokens_per_second": 71850.1744497252,
"prefix_length": 2048
},
"ratios": {
"decode_throughput": 1.0496216169270371,
"persistent_kv": 0.8730769230769231,
"prefill_throughput": 0.8884922352968238
}
},
{
"batch_size": 4,
"dense": {
"batch_size": 4,
"decode_seconds": 0.23466505793233713,
"decode_steps": 32,
"decode_tokens_per_second": 545.458284790347,
"first_cached_token_ms": 7.920100043217341,
"peak_allocated_bytes": 4261779456,
"persistent_kv_bytes": 817889280,
"prefill_seconds": 0.09082256692151229,
"prefill_tokens_per_second": 90197.84705137681,
"prefix_length": 2048
},
"prefix_length": 2048,
"q25": {
"batch_size": 4,
"decode_seconds": 0.2278061956167221,
"decode_steps": 32,
"decode_tokens_per_second": 561.881118524786,
"first_cached_token_ms": 7.969736432035764,
"peak_allocated_bytes": 4070479872,
"persistent_kv_bytes": 714080256,
"prefill_seconds": 0.0923770306011041,
"prefill_tokens_per_second": 88680.05332812775,
"prefix_length": 2048
},
"ratios": {
"decode_throughput": 1.0301083221070724,
"persistent_kv": 0.8730769230769231,
"prefill_throughput": 0.9831726169430127
}
},
{
"batch_size": 8,
"dense": {
"batch_size": 8,
"decode_seconds": 0.25229039539893466,
"decode_steps": 32,
"decode_tokens_per_second": 1014.7037091729136,
"first_cached_token_ms": 8.601184313495954,
"peak_allocated_bytes": 6273153024,
"persistent_kv_bytes": 1635778560,
"prefill_seconds": 0.19029431355496249,
"prefill_tokens_per_second": 86098.2112072825,
"prefix_length": 2048
},
"prefix_length": 2048,
"q25": {
"batch_size": 8,
"decode_seconds": 0.23515919471780458,
"decode_steps": 32,
"decode_tokens_per_second": 1088.6242415790068,
"first_cached_token_ms": 7.94680913289388,
"peak_allocated_bytes": 5870467072,
"persistent_kv_bytes": 1428160512,
"prefill_seconds": 0.19251583764950433,
"prefill_tokens_per_second": 85104.6864509341,
"prefix_length": 2048
},
"ratios": {
"decode_throughput": 1.072849376362629,
"persistent_kv": 0.8730769230769231,
"prefill_throughput": 0.9884605644830823
}
},
{
"batch_size": 16,
"dense": {
"batch_size": 16,
"decode_seconds": 0.3312743529677391,
"decode_steps": 32,
"decode_tokens_per_second": 1545.5467512447628,
"first_cached_token_ms": 11.202741414308548,
"peak_allocated_bytes": 10296591872,
"persistent_kv_bytes": 3271557120,
"prefill_seconds": 0.40776471979916096,
"prefill_tokens_per_second": 80360.06650143603,
"prefix_length": 2048
},
"prefix_length": 2048,
"q25": {
"batch_size": 16,
"decode_seconds": 0.29237379630406696,
"decode_steps": 32,
"decode_tokens_per_second": 1751.1829256665776,
"first_cached_token_ms": 9.928669159611067,
"peak_allocated_bytes": 9478374912,
"persistent_kv_bytes": 2856321024,
"prefill_seconds": 0.41604801764090854,
"prefill_tokens_per_second": 78760.13972089658,
"prefix_length": 2048
},
"ratios": {
"decode_throughput": 1.1330507629459918,
"persistent_kv": 0.8730769230769231,
"prefill_throughput": 0.9800905244334155
}
},
{
"batch_size": 1,
"dense": {
"batch_size": 1,
"decode_seconds": 0.23166167673965296,
"decode_steps": 32,
"decode_tokens_per_second": 138.13247167317354,
"first_cached_token_ms": 7.798002411921819,
"peak_allocated_bytes": 3247719936,
"persistent_kv_bytes": 405798912,
"prefill_seconds": 0.04814621185262998,
"prefill_tokens_per_second": 85074.19052068697,
"prefix_length": 4096
},
"prefix_length": 4096,
"q25": {
"batch_size": 1,
"decode_seconds": 0.22054913143316904,
"decode_steps": 32,
"decode_tokens_per_second": 145.0923873155069,
"first_cached_token_ms": 7.15858240922292,
"peak_allocated_bytes": 3113641472,
"persistent_kv_bytes": 329515008,
"prefill_seconds": 0.05360218447943529,
"prefill_tokens_per_second": 76414.79614644154,
"prefix_length": 4096
},
"ratios": {
"decode_throughput": 1.0503858039896714,
"persistent_kv": 0.812015503875969,
"prefill_throughput": 0.8982136142436785
}
},
{
"batch_size": 4,
"dense": {
"batch_size": 4,
"decode_seconds": 0.25190934414664906,
"decode_steps": 32,
"decode_tokens_per_second": 508.1193015431964,
"first_cached_token_ms": 8.424771949648857,
"peak_allocated_bytes": 6247872512,
"persistent_kv_bytes": 1623195648,
"prefill_seconds": 0.20185994543135166,
"prefill_tokens_per_second": 81165.18591635038,
"prefix_length": 4096
},
"prefix_length": 4096,
"q25": {
"batch_size": 4,
"decode_seconds": 0.231443390250206,
"decode_steps": 32,
"decode_tokens_per_second": 553.0510068212504,
"first_cached_token_ms": 7.94626772403717,
"peak_allocated_bytes": 5648562176,
"persistent_kv_bytes": 1318060032,
"prefill_seconds": 0.20270894095301628,
"prefill_tokens_per_second": 80825.2459066296,
"prefix_length": 4096
},
"ratios": {
"decode_throughput": 1.0884274719373839,
"persistent_kv": 0.812015503875969,
"prefill_throughput": 0.9958117509880267
}
},
{
"batch_size": 8,
"dense": {
"batch_size": 8,
"decode_seconds": 0.321361793205142,
"decode_steps": 32,
"decode_tokens_per_second": 796.6099437234028,
"first_cached_token_ms": 10.821449880798658,
"peak_allocated_bytes": 10243749888,
"persistent_kv_bytes": 3246391296,
"prefill_seconds": 0.4328818339854479,
"prefill_tokens_per_second": 75697.33222184962,
"prefix_length": 4096
},
"prefix_length": 4096,
"q25": {
"batch_size": 8,
"decode_seconds": 0.27276255128284294,
"decode_steps": 32,
"decode_tokens_per_second": 938.5452614224124,
"first_cached_token_ms": 9.249638145168623,
"peak_allocated_bytes": 9035757568,
"persistent_kv_bytes": 2636120064,
"prefill_seconds": 0.43729556041459244,
"prefill_tokens_per_second": 74933.30133270327,
"prefix_length": 4096
},
"ratios": {
"decode_throughput": 1.1781741727144346,
"persistent_kv": 0.812015503875969,
"prefill_throughput": 0.9899067659754881
}
},
{
"batch_size": 16,
"dense": {
"batch_size": 16,
"decode_seconds": 0.4728281206140916,
"decode_steps": 32,
"decode_tokens_per_second": 1082.8459173177632,
"first_cached_token_ms": 15.76905200878779,
"peak_allocated_bytes": 18239866368,
"persistent_kv_bytes": 6492782592,
"prefill_seconds": 0.8857398132483164,
"prefill_tokens_per_second": 73990.12556481644,
"prefix_length": 4096
},
"prefix_length": 4096,
"q25": {
"batch_size": 16,
"decode_seconds": 0.3972581879546245,
"decode_steps": 32,
"decode_tokens_per_second": 1288.8343538899733,
"first_cached_token_ms": 13.64798347155253,
"peak_allocated_bytes": 15803827712,
"persistent_kv_bytes": 5272240128,
"prefill_seconds": 0.8932362391302983,
"prefill_tokens_per_second": 73369.1683443221,
"prefix_length": 4096
},
"ratios": {
"decode_throughput": 1.1902287604153772,
"persistent_kv": 0.812015503875969,
"prefill_throughput": 0.9916075663373436
}
},
{
"batch_size": 1,
"dense": {
"batch_size": 1,
"decode_seconds": 0.22510927046338716,
"decode_steps": 32,
"decode_tokens_per_second": 142.1531860244051,
"first_cached_token_ms": 7.497106368343036,
"peak_allocated_bytes": 4234716672,
"persistent_kv_bytes": 805306368,
"prefill_seconds": 0.10965904593467712,
"prefill_tokens_per_second": 74412.46575189826,
"prefix_length": 8160
},
"prefix_length": 8160,
"q25": {
"batch_size": 1,
"decode_seconds": 0.224536320194602,
"decode_steps": 32,
"decode_tokens_per_second": 142.5159189046392,
"first_cached_token_ms": 7.5872062395016355,
"peak_allocated_bytes": 3902450176,
"persistent_kv_bytes": 629145600,
"prefill_seconds": 0.11838383475939433,
"prefill_tokens_per_second": 68928.32975536353,
"prefix_length": 8160
},
"ratios": {
"decode_throughput": 1.0025517041888306,
"persistent_kv": 0.78125,
"prefill_throughput": 0.9263008429955861
}
},
{
"batch_size": 4,
"dense": {
"batch_size": 4,
"decode_seconds": 0.31844382857282955,
"decode_steps": 32,
"decode_tokens_per_second": 401.95472015789375,
"first_cached_token_ms": 10.655300070842108,
"peak_allocated_bytes": 10188224512,
"persistent_kv_bytes": 3221225472,
"prefill_seconds": 0.47874748582641286,
"prefill_tokens_per_second": 68177.9037307254,
"prefix_length": 8160
},
"prefix_length": 8160,
"q25": {
"batch_size": 4,
"decode_seconds": 0.27023141148189705,
"decode_steps": 32,
"decode_tokens_per_second": 473.6681028236971,
"first_cached_token_ms": 8.966594313581785,
"peak_allocated_bytes": 8791914496,
"persistent_kv_bytes": 2516582400,
"prefill_seconds": 0.4703146008153756,
"prefill_tokens_per_second": 69400.35445085618,
"prefix_length": 8160
},
"ratios": {
"decode_throughput": 1.1784115948125533,
"persistent_kv": 0.78125,
"prefill_throughput": 1.0179303066424417
}
},
{
"batch_size": 8,
"dense": {
"batch_size": 8,
"decode_seconds": 0.45806841490169364,
"decode_steps": 32,
"decode_tokens_per_second": 558.8684826805627,
"first_cached_token_ms": 15.288629879554113,
"peak_allocated_bytes": 18128516096,
"persistent_kv_bytes": 6442450944,
"prefill_seconds": 0.9819728036721548,
"prefill_tokens_per_second": 66478.41952025653,
"prefix_length": 8160
},
"prefix_length": 8160,
"q25": {
"batch_size": 8,
"decode_seconds": 0.3758538272231817,
"decode_steps": 32,
"decode_tokens_per_second": 681.1158526476502,
"first_cached_token_ms": 12.771941100557646,
"peak_allocated_bytes": 15313089536,
"persistent_kv_bytes": 5033164800,
"prefill_seconds": 0.9673846984903017,
"prefill_tokens_per_second": 67480.9102334116,
"prefix_length": 8160
},
"ratios": {
"decode_throughput": 1.2187408554168933,
"persistent_kv": 0.78125,
"prefill_throughput": 1.0150799420381769
}
},
{
"batch_size": 16,
"dense": {
"batch_size": 16,
"decode_seconds": 0.7506915175666412,
"decode_steps": 32,
"decode_tokens_per_second": 682.037811829342,
"first_cached_token_ms": 25.01120666662852,
"peak_allocated_bytes": 34005301248,
"persistent_kv_bytes": 12884901888,
"prefill_seconds": 2.1065259805570045,
"prefill_tokens_per_second": 61978.82257567861,
"prefix_length": 8160
},
"prefix_length": 8160,
"q25": {
"batch_size": 16,
"decode_seconds": 0.606967706233263,
"decode_steps": 32,
"decode_tokens_per_second": 843.5374645833858,
"first_cached_token_ms": 20.8244522412618,
"peak_allocated_bytes": 28370909184,
"persistent_kv_bytes": 10066329600,
"prefill_seconds": 2.119566404571136,
"prefill_tokens_per_second": 61597.50396044655,
"prefix_length": 8160
},
"ratios": {
"decode_throughput": 1.2367898816648804,
"persistent_kv": 0.78125,
"prefill_throughput": 0.9938475982700953
}
}
],
"torch_version": "2.5.1+cu124",
"verification": {
"dense": {
"maximum_absolute_logit_difference": 0.0625,
"maximum_absolute_probability_difference": 0.0015385262668132782,
"mean_absolute_logit_difference": 0.009482888504862785,
"numerically_equivalent": true,
"top_token_exact": true
},
"q25": {
"maximum_absolute_logit_difference": 0.078125,
"maximum_absolute_probability_difference": 0.002120431512594223,
"mean_absolute_logit_difference": 0.011962523683905602,
"numerically_equivalent": true,
"top_token_exact": true
}
}
}