Files
hanzo.ai/lib/data/engine.json
zeekayandHanzo Dev 06c7e66278 Engine board: read the evidence hanzo-bench files; the ROCm 27B lane at ROCmFP4-FAST
lib/data/engine.json is the research payload for every committed run (one
kernel-perf experiment per engine per shape); cells and runs derive from it, so
the board shows what Hanzo Research holds. Each row names its model,
quantization and engine release, and the largest model gets its own paragraph.

Co-authored-by: Hanzo Dev <[email protected]>
2026-09-17 20:25:31 -07:00

2756 lines
84 KiB
JSON

{
"experiments": [
{
"id": "kernel-perf:hanzo-engine/metal/dbc/Qwen3-1.7B-Q4_K_M:prefill-500",
"kind": "kernel-perf",
"subject": "hanzo-engine/metal/dbc/Qwen3-1.7B-Q4_K_M",
"task": "prefill-500",
"metric": "tok/s",
"value": 3345.3144399593316,
"n": 6,
"meta": {
"run": "dbc-metal-qwen3-1p7b-20260721-183614",
"host": "dbc",
"backend": "Metal",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 9.449141007726892,
"cv_pct": 0.26910944050758184,
"std": 9.002556972593903,
"samples": [
3345.935419886321,
3341.3060459649837,
3343.8534753368035,
3342.5671751546984,
3335.878809017342,
3362.345714395846
],
"noisy": false,
"versus": {
"subject": "llama.cpp/metal/dbc/Qwen3-1.7B-Q4_K_M",
"ratio": 0.8386555797036607,
"ci": 0.003914914148774247,
"lo": 0.8347406655548865,
"hi": 0.8425704938524349,
"verdict": "LOSS",
"best": 0.842811451030714
}
},
"git_sha": "a0c6d7429233bca12e0f4996bbcc9f1c8ca42820",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784684174
},
{
"id": "kernel-perf:llama.cpp/metal/dbc/Qwen3-1.7B-Q4_K_M:prefill-500",
"kind": "kernel-perf",
"subject": "llama.cpp/metal/dbc/Qwen3-1.7B-Q4_K_M",
"task": "prefill-500",
"metric": "tok/s",
"value": 3988.9014285714284,
"n": 7,
"meta": {
"run": "dbc-metal-qwen3-1p7b-20260721-183614",
"host": "dbc",
"backend": "Metal",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 14.824925197972181,
"cv_pct": 0.4018410102312195,
"std": 16.029041797698973,
"samples": [
3979.11,
3997.27,
4010.04,
3970.73,
3993.64,
4002.44,
3969.08
],
"noisy": false
},
"git_sha": "81ff7abe50b95fb81cc70a6cdba1eb1a02a48f62",
"git_dirty": false,
"lib_versions": {},
"ts": 1784684174
},
{
"id": "kernel-perf:hanzo-engine/metal/dbc/Qwen3-1.7B-Q4_K_M:prefill-512",
"kind": "kernel-perf",
"subject": "hanzo-engine/metal/dbc/Qwen3-1.7B-Q4_K_M",
"task": "prefill-512",
"metric": "tok/s",
"value": 3386.0265916032026,
"n": 6,
"meta": {
"run": "dbc-metal-qwen3-1p7b-20260721-183614",
"host": "dbc",
"backend": "Metal",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 11.627050040761002,
"cv_pct": 0.32715436991685404,
"std": 11.077533960976586,
"samples": [
3386.0306396092838,
3376.119115074077,
3374.8580837570303,
3397.1731617852893,
3380.5690385091184,
3401.409510884421
],
"noisy": false,
"versus": {
"subject": "llama.cpp/metal/dbc/Qwen3-1.7B-Q4_K_M",
"ratio": 0.8263628723684311,
"ci": 0.00842647534866851,
"lo": 0.8179363970197626,
"hi": 0.8347893477170996,
"verdict": "LOSS",
"best": 0.831828090607096
}
},
"git_sha": "a0c6d7429233bca12e0f4996bbcc9f1c8ca42820",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784684174
},
{
"id": "kernel-perf:llama.cpp/metal/dbc/Qwen3-1.7B-Q4_K_M:prefill-512",
"kind": "kernel-perf",
"subject": "llama.cpp/metal/dbc/Qwen3-1.7B-Q4_K_M",
"task": "prefill-512",
"metric": "tok/s",
"value": 4097.505714285714,
"n": 7,
"meta": {
"run": "dbc-metal-qwen3-1p7b-20260721-183614",
"host": "dbc",
"backend": "Metal",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 39.34222431556452,
"cv_pct": 1.0381363380990578,
"std": 42.53769577568535,
"samples": [
4101.51,
4109.72,
4125.36,
4003.03,
4106.52,
4123.31,
4113.09
],
"noisy": false
},
"git_sha": "81ff7abe50b95fb81cc70a6cdba1eb1a02a48f62",
"git_dirty": false,
"lib_versions": {},
"ts": 1784684174
},
{
"id": "kernel-perf:hanzo-engine/metal/dbc/Qwen3-1.7B-Q4_K_M:prefill-2048",
"kind": "kernel-perf",
"subject": "hanzo-engine/metal/dbc/Qwen3-1.7B-Q4_K_M",
"task": "prefill-2048",
"metric": "tok/s",
"value": 3301.4919185562653,
"n": 6,
"meta": {
"run": "dbc-metal-qwen3-1p7b-20260721-183614",
"host": "dbc",
"backend": "Metal",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 4.322521222057199,
"cv_pct": 0.12473847900653433,
"std": 4.118230803730734,
"samples": [
3300.5137476880914,
3301.418013845644,
3299.0095923778063,
3297.246006185015,
3301.549290372121,
3309.2148608689163
],
"noisy": false,
"versus": {
"subject": "llama.cpp/metal/dbc/Qwen3-1.7B-Q4_K_M",
"ratio": 0.8930536915485686,
"ci": 0.026943387835982095,
"lo": 0.8661103037125866,
"hi": 0.9199970793845507,
"verdict": "LOSS",
"best": 0.8757997271368323
}
},
"git_sha": "a0c6d7429233bca12e0f4996bbcc9f1c8ca42820",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784684174
},
{
"id": "kernel-perf:llama.cpp/metal/dbc/Qwen3-1.7B-Q4_K_M:prefill-2048",
"kind": "kernel-perf",
"subject": "llama.cpp/metal/dbc/Qwen3-1.7B-Q4_K_M",
"task": "prefill-2048",
"metric": "tok/s",
"value": 3696.8571428571427,
"n": 7,
"meta": {
"run": "dbc-metal-qwen3-1p7b-20260721-183614",
"host": "dbc",
"backend": "Metal",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 111.42893411448485,
"cv_pct": 3.2589699680486492,
"std": 120.47946404737563,
"samples": [
3835.42,
3830.96,
3725.68,
3535.99,
3551.87,
3666.11,
3731.97
],
"noisy": false
},
"git_sha": "81ff7abe50b95fb81cc70a6cdba1eb1a02a48f62",
"git_dirty": false,
"lib_versions": {},
"ts": 1784684174
},
{
"id": "kernel-perf:hanzo-engine/metal/dbc/Qwen3-1.7B-Q4_K_M:prefill-4096",
"kind": "kernel-perf",
"subject": "hanzo-engine/metal/dbc/Qwen3-1.7B-Q4_K_M",
"task": "prefill-4096",
"metric": "tok/s",
"value": 2767.252037083752,
"n": 6,
"meta": {
"run": "dbc-metal-qwen3-1p7b-20260721-183614",
"host": "dbc",
"backend": "Metal",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 113.56929907121597,
"cv_pct": 3.910081215541044,
"std": 108.20180208868867,
"samples": [
2912.374311201878,
2840.2869093909153,
2798.2423540213804,
2748.663847392938,
2698.1182179351017,
2605.826582560297
],
"noisy": false,
"versus": {
"subject": "llama.cpp/metal/dbc/Qwen3-1.7B-Q4_K_M",
"ratio": 0.9665553247306283,
"ci": 0.0460698472564181,
"lo": 0.9204854774742103,
"hi": 1.0126251719870465,
"verdict": "PARITY",
"best": 0.9936351082049237
}
},
"git_sha": "a0c6d7429233bca12e0f4996bbcc9f1c8ca42820",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784684174
},
{
"id": "kernel-perf:llama.cpp/metal/dbc/Qwen3-1.7B-Q4_K_M:prefill-4096",
"kind": "kernel-perf",
"subject": "llama.cpp/metal/dbc/Qwen3-1.7B-Q4_K_M",
"task": "prefill-4096",
"metric": "tok/s",
"value": 2863.0042857142857,
"n": 7,
"meta": {
"run": "dbc-metal-qwen3-1p7b-20260721-183614",
"host": "dbc",
"backend": "Metal",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 69.39658716207623,
"cv_pct": 2.620783618255737,
"std": 75.03314730995967,
"samples": [
2931.03,
2928.25,
2914.41,
2883.41,
2866.82,
2759.68,
2757.43
],
"noisy": false
},
"git_sha": "81ff7abe50b95fb81cc70a6cdba1eb1a02a48f62",
"git_dirty": false,
"lib_versions": {},
"ts": 1784684174
},
{
"id": "kernel-perf:hanzo-engine/metal/dbc/Qwen3-1.7B-Q4_K_M:decode-128",
"kind": "kernel-perf",
"subject": "hanzo-engine/metal/dbc/Qwen3-1.7B-Q4_K_M",
"task": "decode-128",
"metric": "tok/s",
"value": 233.4362181514575,
"n": 6,
"meta": {
"run": "dbc-metal-qwen3-1p7b-20260721-183614",
"host": "dbc",
"backend": "Metal",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 1.460233840759422,
"cv_pct": 0.5959745560558302,
"std": 1.391220464801668,
"samples": [
232.75441019207736,
234.56848379076735,
235.34282521845728,
233.75357096311524,
232.65509676778558,
231.5429219765422
],
"noisy": false,
"versus": {
"subject": "llama.cpp/metal/dbc/Qwen3-1.7B-Q4_K_M",
"ratio": 0.9198048361124729,
"ci": 0.006785074017174069,
"lo": 0.9130197620952988,
"hi": 0.926589910129647,
"verdict": "LOSS",
"best": 0.9212439627751183
}
},
"git_sha": "a0c6d7429233bca12e0f4996bbcc9f1c8ca42820",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784684174
},
{
"id": "kernel-perf:llama.cpp/metal/dbc/Qwen3-1.7B-Q4_K_M:decode-128",
"kind": "kernel-perf",
"subject": "llama.cpp/metal/dbc/Qwen3-1.7B-Q4_K_M",
"task": "decode-128",
"metric": "tok/s",
"value": 253.78885714285713,
"n": 7,
"meta": {
"run": "dbc-metal-qwen3-1p7b-20260721-183614",
"host": "dbc",
"backend": "Metal",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 0.9922148462280578,
"cv_pct": 0.4227155475919854,
"std": 1.07280495719887,
"samples": [
253.753,
255.462,
254.273,
254.594,
252.513,
253.304,
252.623
],
"noisy": false
},
"git_sha": "81ff7abe50b95fb81cc70a6cdba1eb1a02a48f62",
"git_dirty": false,
"lib_versions": {},
"ts": 1784684174
},
{
"id": "kernel-perf:hanzo-engine/metal/dbc/Qwen3-4B-Q4_K_M:prefill-512",
"kind": "kernel-perf",
"subject": "hanzo-engine/metal/dbc/Qwen3-4B-Q4_K_M",
"task": "prefill-512",
"metric": "tok/s",
"value": 1411.013702479241,
"n": 6,
"meta": {
"run": "dbc-metal-qwen3-4b-20260721-181831",
"host": "dbc",
"backend": "Metal",
"model": "Qwen3-4B-Q4_K_M",
"model_sha256": "7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5",
"model_bytes": 2497280256,
"gpu": "",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 7.515758369065985,
"cv_pct": 0.5074755540888193,
"std": 7.160549604925692,
"samples": [
1416.810952858524,
1404.8514949456915,
1400.2849114856524,
1410.3558058022584,
1417.202636141945,
1416.5764136413752
],
"noisy": false,
"versus": {
"subject": "llama.cpp/metal/dbc/Qwen3-4B-Q4_K_M",
"ratio": 0.8664712074470609,
"ci": 0.012660370128003072,
"lo": 0.8538108373190578,
"hi": 0.8791315775750639,
"verdict": "LOSS",
"best": 0.8628344492409096
}
},
"git_sha": "a0c6d7429233bca12e0f4996bbcc9f1c8ca42820",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784683111
},
{
"id": "kernel-perf:llama.cpp/metal/dbc/Qwen3-4B-Q4_K_M:prefill-512",
"kind": "kernel-perf",
"subject": "llama.cpp/metal/dbc/Qwen3-4B-Q4_K_M",
"task": "prefill-512",
"metric": "tok/s",
"value": 1628.4600000000003,
"n": 7,
"meta": {
"run": "dbc-metal-qwen3-4b-20260721-181831",
"host": "dbc",
"backend": "Metal",
"model": "Qwen3-4B-Q4_K_M",
"model_sha256": "7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5",
"model_bytes": 2497280256,
"gpu": "",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 22.156745800350034,
"cv_pct": 1.47110587672995,
"std": 23.956370760196553,
"samples": [
1652.53,
1646.56,
1640.88,
1645.76,
1617.88,
1604.57,
1591.04
],
"noisy": false
},
"git_sha": "81ff7abe50b95fb81cc70a6cdba1eb1a02a48f62",
"git_dirty": false,
"lib_versions": {},
"ts": 1784683111
},
{
"id": "kernel-perf:hanzo-engine/metal/dbc/Qwen3-4B-Q4_K_M:prefill-2048",
"kind": "kernel-perf",
"subject": "hanzo-engine/metal/dbc/Qwen3-4B-Q4_K_M",
"task": "prefill-2048",
"metric": "tok/s",
"value": 1221.3693781111392,
"n": 6,
"meta": {
"run": "dbc-metal-qwen3-4b-20260721-181831",
"host": "dbc",
"backend": "Metal",
"model": "Qwen3-4B-Q4_K_M",
"model_sha256": "7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5",
"model_bytes": 2497280256,
"gpu": "",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 48.86540861818514,
"cv_pct": 3.811782009890618,
"std": 46.55593822915333,
"samples": [
1284.435853004101,
1258.1032051607417,
1237.24749284452,
1205.7699743553096,
1173.2222614487864,
1169.4374818533768
],
"noisy": false,
"versus": {
"subject": "llama.cpp/metal/dbc/Qwen3-4B-Q4_K_M",
"ratio": 0.9980418341953848,
"ci": 0.044312998075287734,
"lo": 0.9537288361200971,
"hi": 1.0423548322706726,
"verdict": "PARITY",
"best": 1.0508590600223116
}
},
"git_sha": "a0c6d7429233bca12e0f4996bbcc9f1c8ca42820",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784683111
},
{
"id": "kernel-perf:llama.cpp/metal/dbc/Qwen3-4B-Q4_K_M:prefill-2048",
"kind": "kernel-perf",
"subject": "llama.cpp/metal/dbc/Qwen3-4B-Q4_K_M",
"task": "prefill-2048",
"metric": "tok/s",
"value": 1223.765714285714,
"n": 7,
"meta": {
"run": "dbc-metal-qwen3-4b-20260721-181831",
"host": "dbc",
"backend": "Metal",
"model": "Qwen3-4B-Q4_K_M",
"model_sha256": "7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5",
"model_bytes": 2497280256,
"gpu": "",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 23.560530017024384,
"cv_pct": 2.08162178912424,
"std": 25.474173756403314,
"samples": [
1278.1,
1232.8,
1211.57,
1205.92,
1211.81,
1209.82,
1216.34
],
"noisy": false
},
"git_sha": "81ff7abe50b95fb81cc70a6cdba1eb1a02a48f62",
"git_dirty": false,
"lib_versions": {},
"ts": 1784683111
},
{
"id": "kernel-perf:hanzo-engine/metal/dbc/Qwen3-4B-Q4_K_M:decode-128",
"kind": "kernel-perf",
"subject": "hanzo-engine/metal/dbc/Qwen3-4B-Q4_K_M",
"task": "decode-128",
"metric": "tok/s",
"value": 128.1940163302488,
"n": 6,
"meta": {
"run": "dbc-metal-qwen3-4b-20260721-181831",
"host": "dbc",
"backend": "Metal",
"model": "Qwen3-4B-Q4_K_M",
"model_sha256": "7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5",
"model_bytes": 2497280256,
"gpu": "",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 0.7999498020560175,
"cv_pct": 0.5945228206477813,
"std": 0.7621426817882724,
"samples": [
128.99582337932492,
128.74412354924343,
127.8748450106534,
127.67216727723167,
127.08589951994583,
128.79123924509364
],
"noisy": false,
"versus": {
"subject": "llama.cpp/metal/dbc/Qwen3-4B-Q4_K_M",
"ratio": 0.9383401190094804,
"ci": 0.015466515973908414,
"lo": 0.922873603035572,
"hi": 0.9538066349833889,
"verdict": "LOSS",
"best": 0.9298671715936199
}
},
"git_sha": "a0c6d7429233bca12e0f4996bbcc9f1c8ca42820",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784683111
},
{
"id": "kernel-perf:llama.cpp/metal/dbc/Qwen3-4B-Q4_K_M:decode-128",
"kind": "kernel-perf",
"subject": "llama.cpp/metal/dbc/Qwen3-4B-Q4_K_M",
"task": "decode-128",
"metric": "tok/s",
"value": 136.61785714285716,
"n": 7,
"meta": {
"run": "dbc-metal-qwen3-4b-20260721-181831",
"host": "dbc",
"backend": "Metal",
"model": "Qwen3-4B-Q4_K_M",
"model_sha256": "7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5",
"model_bytes": 2497280256,
"gpu": "",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 2.084238550968209,
"cv_pct": 1.6495102024554842,
"std": 2.2535254919474874,
"samples": [
138.312,
138.725,
138.108,
137.841,
136.217,
134.106,
133.016
],
"noisy": false
},
"git_sha": "81ff7abe50b95fb81cc70a6cdba1eb1a02a48f62",
"git_dirty": false,
"lib_versions": {},
"ts": 1784683111
},
{
"id": "kernel-perf:hanzo-engine/rocm/evo/Qwen3-1.7B-Q4_K_M:prefill-500",
"kind": "kernel-perf",
"subject": "hanzo-engine/rocm/evo/Qwen3-1.7B-Q4_K_M",
"task": "prefill-500",
"metric": "tok/s",
"value": 3603.7696463334705,
"n": 6,
"meta": {
"run": "evo-rocm-qwen3-1p7b-20260721-174808",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 19.428680845263067,
"cv_pct": 0.5136411855697941,
"std": 18.510445136631613,
"samples": [
3636.2007610146397,
3612.3648184450008,
3600.1504690088827,
3586.791645504297,
3587.330001005098,
3599.780183022904
],
"noisy": false,
"versus": {
"subject": "llama.cpp/rocm/evo/Qwen3-1.7B-Q4_K_M",
"ratio": 0.7703607601594777,
"ci": 0.030966066163433396,
"lo": 0.7393946939960443,
"hi": 0.8013268263229112,
"verdict": "LOSS",
"best": 0.7616356166508818
}
},
"git_sha": "1ac219c43d04a69ef4f224e037bef309a4557728",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784681288
},
{
"id": "kernel-perf:llama.cpp/rocm/evo/Qwen3-1.7B-Q4_K_M:prefill-500",
"kind": "kernel-perf",
"subject": "llama.cpp/rocm/evo/Qwen3-1.7B-Q4_K_M",
"task": "prefill-500",
"metric": "tok/s",
"value": 4678.028571428572,
"n": 7,
"meta": {
"run": "evo-rocm-qwen3-1p7b-20260721-174808",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 186.34300613862197,
"cv_pct": 4.306904858041425,
"std": 201.47823980342304,
"samples": [
4222.06,
4741.07,
4748.23,
4758.55,
4737.71,
4764.38,
4774.2
],
"noisy": false
},
"git_sha": "0821c5fcfd729af70037bc1e9e60769d42c081ba",
"git_dirty": false,
"lib_versions": {},
"ts": 1784681288
},
{
"id": "kernel-perf:hanzo-engine/rocm/evo/Qwen3-1.7B-Q4_K_M:prefill-512",
"kind": "kernel-perf",
"subject": "hanzo-engine/rocm/evo/Qwen3-1.7B-Q4_K_M",
"task": "prefill-512",
"metric": "tok/s",
"value": 3693.5574620386797,
"n": 6,
"meta": {
"run": "evo-rocm-qwen3-1p7b-20260721-174808",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 14.034596918897545,
"cv_pct": 0.36201672287079095,
"std": 13.371295681421985,
"samples": [
3716.309557988892,
3678.3342134753734,
3694.267901049525,
3685.72165324974,
3686.9305168563874,
3699.7809296121586
],
"noisy": false,
"versus": {
"subject": "llama.cpp/rocm/evo/Qwen3-1.7B-Q4_K_M",
"ratio": 0.7788233422648816,
"ci": 0.035267136645817376,
"lo": 0.7435562056190642,
"hi": 0.814090478910699,
"verdict": "LOSS",
"best": 0.7811058609952131
}
},
"git_sha": "1ac219c43d04a69ef4f224e037bef309a4557728",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784681288
},
{
"id": "kernel-perf:llama.cpp/rocm/evo/Qwen3-1.7B-Q4_K_M:prefill-512",
"kind": "kernel-perf",
"subject": "llama.cpp/rocm/evo/Qwen3-1.7B-Q4_K_M",
"task": "prefill-512",
"metric": "tok/s",
"value": 4742.484285714286,
"n": 7,
"meta": {
"run": "evo-rocm-qwen3-1p7b-20260721-174808",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 213.99456958833997,
"cv_pct": 4.878787504002017,
"std": 231.3757307106879,
"samples": [
4222.45,
4811.28,
4859.19,
4870.98,
4845.32,
4792.08,
4796.09
],
"noisy": false
},
"git_sha": "0821c5fcfd729af70037bc1e9e60769d42c081ba",
"git_dirty": false,
"lib_versions": {},
"ts": 1784681288
},
{
"id": "kernel-perf:hanzo-engine/rocm/evo/Qwen3-1.7B-Q4_K_M:prefill-2048",
"kind": "kernel-perf",
"subject": "hanzo-engine/rocm/evo/Qwen3-1.7B-Q4_K_M",
"task": "prefill-2048",
"metric": "tok/s",
"value": 3506.1271431747396,
"n": 6,
"meta": {
"run": "evo-rocm-qwen3-1p7b-20260721-174808",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 47.03142896330513,
"cv_pct": 1.278009440188284,
"std": 44.80863587477697,
"samples": [
3433.7760155248798,
3489.2319974970483,
3524.244230255939,
3570.2949666652735,
3504.6596469292244,
3514.5560021760725
],
"noisy": false,
"versus": {
"subject": "llama.cpp/rocm/evo/Qwen3-1.7B-Q4_K_M",
"ratio": 0.7703502798460101,
"ci": 0.010880584700293678,
"lo": 0.7594696951457165,
"hi": 0.7812308645463037,
"verdict": "LOSS",
"best": 0.7804076947230161
}
},
"git_sha": "1ac219c43d04a69ef4f224e037bef309a4557728",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784681288
},
{
"id": "kernel-perf:llama.cpp/rocm/evo/Qwen3-1.7B-Q4_K_M:prefill-2048",
"kind": "kernel-perf",
"subject": "llama.cpp/rocm/evo/Qwen3-1.7B-Q4_K_M",
"task": "prefill-2048",
"metric": "tok/s",
"value": 4551.341428571429,
"n": 7,
"meta": {
"run": "evo-rocm-qwen3-1p7b-20260721-174808",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 20.126975341065364,
"cv_pct": 0.4781389779865493,
"std": 21.761737391249845,
"samples": [
4531.45,
4556.29,
4563.72,
4574.91,
4517.11,
4542.9,
4573.01
],
"noisy": false
},
"git_sha": "0821c5fcfd729af70037bc1e9e60769d42c081ba",
"git_dirty": false,
"lib_versions": {},
"ts": 1784681288
},
{
"id": "kernel-perf:hanzo-engine/rocm/evo/Qwen3-1.7B-Q4_K_M:prefill-4096",
"kind": "kernel-perf",
"subject": "hanzo-engine/rocm/evo/Qwen3-1.7B-Q4_K_M",
"task": "prefill-4096",
"metric": "tok/s",
"value": 2840.8166152090694,
"n": 6,
"meta": {
"run": "evo-rocm-qwen3-1p7b-20260721-174808",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 93.02850751716284,
"cv_pct": 3.1199411533969688,
"std": 88.63180667044657,
"samples": [
2671.6021838453726,
2865.970602359723,
2832.0026504668554,
2927.6965684156166,
2884.0190520129254,
2863.608634153923
],
"noisy": false,
"versus": {
"subject": "llama.cpp/rocm/evo/Qwen3-1.7B-Q4_K_M",
"ratio": 0.6923304080150585,
"ci": 0.022831111750039986,
"lo": 0.6694992962650186,
"hi": 0.7151615197650985,
"verdict": "LOSS",
"best": 0.710585266549425
}
},
"git_sha": "1ac219c43d04a69ef4f224e037bef309a4557728",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784681288
},
{
"id": "kernel-perf:llama.cpp/rocm/evo/Qwen3-1.7B-Q4_K_M:prefill-4096",
"kind": "kernel-perf",
"subject": "llama.cpp/rocm/evo/Qwen3-1.7B-Q4_K_M",
"task": "prefill-4096",
"metric": "tok/s",
"value": 4103.267142857143,
"n": 7,
"meta": {
"run": "evo-rocm-qwen3-1p7b-20260721-174808",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 15.956642285870466,
"cv_pct": 0.4204620149842528,
"std": 17.25267970904392,
"samples": [
4102.0,
4069.92,
4106.74,
4115.89,
4114.48,
4120.12,
4093.72
],
"noisy": false
},
"git_sha": "0821c5fcfd729af70037bc1e9e60769d42c081ba",
"git_dirty": false,
"lib_versions": {},
"ts": 1784681288
},
{
"id": "kernel-perf:hanzo-engine/rocm/evo/Qwen3-1.7B-Q4_K_M:decode-128",
"kind": "kernel-perf",
"subject": "hanzo-engine/rocm/evo/Qwen3-1.7B-Q4_K_M",
"task": "decode-128",
"metric": "tok/s",
"value": 113.71188888032852,
"n": 6,
"meta": {
"run": "evo-rocm-qwen3-1p7b-20260721-174808",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 1.0700571270017136,
"cv_pct": 0.8965502557414742,
"std": 1.0194842305650462,
"samples": [
114.99943831376704,
115.0052566140861,
113.42389994774543,
112.95201983100925,
112.82149845171948,
113.06922012364383
],
"noisy": false,
"versus": {
"subject": "llama.cpp/rocm/evo/Qwen3-1.7B-Q4_K_M",
"ratio": 0.8501351296507957,
"ci": 0.009135294553051664,
"lo": 0.8409998350977441,
"hi": 0.8592704242038474,
"verdict": "LOSS",
"best": 0.8565670111206072
}
},
"git_sha": "1ac219c43d04a69ef4f224e037bef309a4557728",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784681288
},
{
"id": "kernel-perf:llama.cpp/rocm/evo/Qwen3-1.7B-Q4_K_M:decode-128",
"kind": "kernel-perf",
"subject": "llama.cpp/rocm/evo/Qwen3-1.7B-Q4_K_M",
"task": "decode-128",
"metric": "tok/s",
"value": 133.75742857142862,
"n": 7,
"meta": {
"run": "evo-rocm-qwen3-1p7b-20260721-174808",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 0.6939578386594808,
"cv_pct": 0.5609578441326822,
"std": 0.7503227876815983,
"samples": [
132.739,
134.139,
134.263,
134.254,
134.223,
134.094,
132.59
],
"noisy": false
},
"git_sha": "0821c5fcfd729af70037bc1e9e60769d42c081ba",
"git_dirty": false,
"lib_versions": {},
"ts": 1784681288
},
{
"id": "kernel-perf:hanzo-engine/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST:prefill-500",
"kind": "kernel-perf",
"subject": "hanzo-engine/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST",
"task": "prefill-500",
"metric": "tok/s",
"value": 50.85862978289208,
"n": 2,
"meta": {
"run": "evo-rocm-qwen3p8-27b-rocmfp4-20260917-122324",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3.8-27B-ROCmFP4-FAST",
"model_sha256": "4f97e8db4c44a6bfefe4244be1424c4f4b996de1b9737801b87a061dd5a7692a",
"model_bytes": 14562235776,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": "greedy-argmax",
"reps": 3,
"concurrency": 1,
"args": "--paged-attn --pa-ctxt-len 8192",
"ci95": 0.6631967370325226,
"cv_pct": 0.14513891503523088,
"std": 0.07381566346867437,
"samples": [
50.91082533908856,
50.80643422669559
],
"noisy": false,
"versus": {
"subject": "llama.cpp/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST",
"ratio": 0.13536371603610525,
"ci": 0.007619611238065792,
"lo": 0.12774410479803947,
"hi": 0.14298332727417104,
"verdict": "LOSS",
"best": 0.13568106291917953
}
},
"git_sha": "e619ff15a036725e64ab3b6e17225f117f31af39",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.95",
"hanzo-metal-kernels": "0.11.45",
"hanzo-ml": "0.11.98",
"hanzo-rocm-kernels": "0.11.38"
},
"ts": 1789673010
},
{
"id": "kernel-perf:llama.cpp/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST:prefill-500",
"kind": "kernel-perf",
"subject": "llama.cpp/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST",
"task": "prefill-500",
"metric": "tok/s",
"value": 375.7183333333334,
"n": 3,
"meta": {
"run": "evo-rocm-qwen3p8-27b-rocmfp4-20260917-122324",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3.8-27B-ROCmFP4-FAST",
"model_sha256": "4f97e8db4c44a6bfefe4244be1424c4f4b996de1b9737801b87a061dd5a7692a",
"model_bytes": 14562235776,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": "greedy-argmax",
"reps": 3,
"concurrency": 1,
"ci95": 20.573836893103767,
"cv_pct": 2.2041552791363594,
"std": 8.281415478849812,
"samples": [
366.158,
380.32,
380.677
],
"noisy": false
},
"git_sha": "f00746ff0c7cdf55e04af90d3583c3646346bcf3",
"git_dirty": false,
"lib_versions": {},
"ts": 1789673010
},
{
"id": "kernel-perf:hanzo-engine/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST:prefill-512",
"kind": "kernel-perf",
"subject": "hanzo-engine/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST",
"task": "prefill-512",
"metric": "tok/s",
"value": 50.37892672714605,
"n": 2,
"meta": {
"run": "evo-rocm-qwen3p8-27b-rocmfp4-20260917-122324",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3.8-27B-ROCmFP4-FAST",
"model_sha256": "4f97e8db4c44a6bfefe4244be1424c4f4b996de1b9737801b87a061dd5a7692a",
"model_bytes": 14562235776,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": "greedy-argmax",
"reps": 3,
"concurrency": 1,
"args": "--paged-attn --pa-ctxt-len 8192",
"ci95": 2.7697102639172573,
"cv_pct": 0.6119156719721572,
"std": 0.3082765480147765,
"samples": [
50.596911164728084,
50.16094228956403
],
"noisy": false,
"versus": {
"subject": "llama.cpp/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST",
"ratio": 0.14654636194002868,
"ci": 0.011943538614464816,
"lo": 0.13460282332556386,
"hi": 0.1584899005544935,
"verdict": "LOSS",
"best": 0.14446139747808054
}
},
"git_sha": "e619ff15a036725e64ab3b6e17225f117f31af39",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.95",
"hanzo-metal-kernels": "0.11.45",
"hanzo-ml": "0.11.98",
"hanzo-rocm-kernels": "0.11.38"
},
"ts": 1789673010
},
{
"id": "kernel-perf:llama.cpp/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST:prefill-512",
"kind": "kernel-perf",
"subject": "llama.cpp/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST",
"task": "prefill-512",
"metric": "tok/s",
"value": 343.7746666666667,
"n": 3,
"meta": {
"run": "evo-rocm-qwen3p8-27b-rocmfp4-20260917-122324",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3.8-27B-ROCmFP4-FAST",
"model_sha256": "4f97e8db4c44a6bfefe4244be1424c4f4b996de1b9737801b87a061dd5a7692a",
"model_bytes": 14562235776,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": "greedy-argmax",
"reps": 3,
"concurrency": 1,
"ci95": 20.68292388610243,
"cv_pct": 2.421739053587674,
"std": 8.325325359007515,
"samples": [
352.281,
343.4,
335.643
],
"noisy": false
},
"git_sha": "f00746ff0c7cdf55e04af90d3583c3646346bcf3",
"git_dirty": false,
"lib_versions": {},
"ts": 1789673010
},
{
"id": "kernel-perf:hanzo-engine/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST:prefill-2048",
"kind": "kernel-perf",
"subject": "hanzo-engine/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST",
"task": "prefill-2048",
"metric": "tok/s",
"value": 49.39610715349064,
"n": 2,
"meta": {
"run": "evo-rocm-qwen3p8-27b-rocmfp4-20260917-122324",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3.8-27B-ROCmFP4-FAST",
"model_sha256": "4f97e8db4c44a6bfefe4244be1424c4f4b996de1b9737801b87a061dd5a7692a",
"model_bytes": 14562235776,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": "greedy-argmax",
"reps": 3,
"concurrency": 1,
"args": "--paged-attn --pa-ctxt-len 8192",
"ci95": 0.5206684420747762,
"cv_pct": 0.11732070889498887,
"std": 0.057951863079003525,
"samples": [
49.35512899812508,
49.4370853088562
],
"noisy": false,
"versus": {
"subject": "llama.cpp/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST",
"ratio": 0.141253616416138,
"ci": 0.0024480418341480104,
"lo": 0.13880557458199,
"hi": 0.14370165825028602,
"verdict": "LOSS",
"best": 0.14083917436956567
}
},
"git_sha": "e619ff15a036725e64ab3b6e17225f117f31af39",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.95",
"hanzo-metal-kernels": "0.11.45",
"hanzo-ml": "0.11.98",
"hanzo-rocm-kernels": "0.11.38"
},
"ts": 1789673010
},
{
"id": "kernel-perf:llama.cpp/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST:prefill-2048",
"kind": "kernel-perf",
"subject": "llama.cpp/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST",
"task": "prefill-2048",
"metric": "tok/s",
"value": 349.69800000000004,
"n": 3,
"meta": {
"run": "evo-rocm-qwen3p8-27b-rocmfp4-20260917-122324",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3.8-27B-ROCmFP4-FAST",
"model_sha256": "4f97e8db4c44a6bfefe4244be1424c4f4b996de1b9737801b87a061dd5a7692a",
"model_bytes": 14562235776,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": "greedy-argmax",
"reps": 3,
"concurrency": 1,
"ci95": 4.810752075580955,
"cv_pct": 0.5537440958317202,
"std": 1.936432028241609,
"samples": [
347.475,
351.018,
350.601
],
"noisy": false
},
"git_sha": "f00746ff0c7cdf55e04af90d3583c3646346bcf3",
"git_dirty": false,
"lib_versions": {},
"ts": 1789673010
},
{
"id": "kernel-perf:hanzo-engine/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST:prefill-4096",
"kind": "kernel-perf",
"subject": "hanzo-engine/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST",
"task": "prefill-4096",
"metric": "tok/s",
"value": 46.91212202128851,
"n": 2,
"meta": {
"run": "evo-rocm-qwen3p8-27b-rocmfp4-20260917-122324",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3.8-27B-ROCmFP4-FAST",
"model_sha256": "4f97e8db4c44a6bfefe4244be1424c4f4b996de1b9737801b87a061dd5a7692a",
"model_bytes": 14562235776,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": "greedy-argmax",
"reps": 3,
"concurrency": 1,
"args": "--paged-attn --pa-ctxt-len 8192",
"ci95": 4.36756009294875,
"cv_pct": 1.036239057959789,
"std": 0.4861217313023468,
"samples": [
47.25586199397454,
46.56838204860247
],
"noisy": false,
"versus": {
"subject": "llama.cpp/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST",
"ratio": 0.13881146744480838,
"ci": 0.012960521884985705,
"lo": 0.12585094555982268,
"hi": 0.15177198932979408,
"verdict": "LOSS",
"best": 0.13947139215229964
}
},
"git_sha": "e619ff15a036725e64ab3b6e17225f117f31af39",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.95",
"hanzo-metal-kernels": "0.11.45",
"hanzo-ml": "0.11.98",
"hanzo-rocm-kernels": "0.11.38"
},
"ts": 1789673010
},
{
"id": "kernel-perf:llama.cpp/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST:prefill-4096",
"kind": "kernel-perf",
"subject": "llama.cpp/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST",
"task": "prefill-4096",
"metric": "tok/s",
"value": 337.9556666666667,
"n": 3,
"meta": {
"run": "evo-rocm-qwen3p8-27b-rocmfp4-20260917-122324",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3.8-27B-ROCmFP4-FAST",
"model_sha256": "4f97e8db4c44a6bfefe4244be1424c4f4b996de1b9737801b87a061dd5a7692a",
"model_bytes": 14562235776,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": "greedy-argmax",
"reps": 3,
"concurrency": 1,
"ci95": 2.3842520515231653,
"cv_pct": 0.28397604959700684,
"std": 0.9597131515892285,
"samples": [
338.995,
337.769,
337.103
],
"noisy": false
},
"git_sha": "f00746ff0c7cdf55e04af90d3583c3646346bcf3",
"git_dirty": false,
"lib_versions": {},
"ts": 1789673010
},
{
"id": "kernel-perf:hanzo-engine/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST:decode-128",
"kind": "kernel-perf",
"subject": "hanzo-engine/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST",
"task": "decode-128",
"metric": "tok/s",
"value": 8.631487685177678,
"n": 2,
"meta": {
"run": "evo-rocm-qwen3p8-27b-rocmfp4-20260917-122324",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3.8-27B-ROCmFP4-FAST",
"model_sha256": "4f97e8db4c44a6bfefe4244be1424c4f4b996de1b9737801b87a061dd5a7692a",
"model_bytes": 14562235776,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": "greedy-argmax",
"reps": 3,
"concurrency": 1,
"args": "--paged-attn --pa-ctxt-len 8192",
"ci95": 1.8024781469622977,
"cv_pct": 2.3242909952325053,
"std": 0.20062089102118735,
"samples": [
8.48962729268891,
8.773348077666448
],
"noisy": false,
"versus": {
"subject": "llama.cpp/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST",
"ratio": 0.6646712781495351,
"ci": 0.13905968753265394,
"lo": 0.5256115906168811,
"hi": 0.803730965682189,
"verdict": "LOSS",
"best": 0.6733242831999054
}
},
"git_sha": "e619ff15a036725e64ab3b6e17225f117f31af39",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.95",
"hanzo-metal-kernels": "0.11.45",
"hanzo-ml": "0.11.98",
"hanzo-rocm-kernels": "0.11.38"
},
"ts": 1789673010
},
{
"id": "kernel-perf:llama.cpp/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST:decode-128",
"kind": "kernel-perf",
"subject": "llama.cpp/rocm/evo/Qwen3.8-27B-ROCmFP4-FAST",
"task": "decode-128",
"metric": "tok/s",
"value": 12.9861,
"n": 3,
"meta": {
"run": "evo-rocm-qwen3p8-27b-rocmfp4-20260917-122324",
"host": "evo",
"backend": "ROCm",
"model": "Qwen3.8-27B-ROCmFP4-FAST",
"model_sha256": "4f97e8db4c44a6bfefe4244be1424c4f4b996de1b9737801b87a061dd5a7692a",
"model_bytes": 14562235776,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": "greedy-argmax",
"reps": 3,
"concurrency": 1,
"ci95": 0.16577890391277134,
"cv_pct": 0.5138540718566573,
"std": 0.06672960362537739,
"samples": [
13.0191,
13.0299,
12.9093
],
"noisy": false
},
"git_sha": "f00746ff0c7cdf55e04af90d3583c3646346bcf3",
"git_dirty": false,
"lib_versions": {},
"ts": 1789673010
},
{
"id": "kernel-perf:hanzo-engine/vulkan/evo/Qwen3-1.7B-Q4_K_M:prefill-500",
"kind": "kernel-perf",
"subject": "hanzo-engine/vulkan/evo/Qwen3-1.7B-Q4_K_M",
"task": "prefill-500",
"metric": "tok/s",
"value": 1098.868949317258,
"n": 6,
"meta": {
"run": "evo-vulkan-qwen3-1p7b-20260721-175906",
"host": "evo",
"backend": "Vulkan",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 241.38544033885984,
"cv_pct": 20.928529664194624,
"std": 229.97711402848614,
"samples": [
1034.2484164746606,
1215.0801901284576,
1465.920292372525,
1082.5891875079808,
1021.839071627322,
773.5365377926018
],
"noisy": true,
"versus": {
"subject": "llama.cpp/vulkan/evo/Qwen3-1.7B-Q4_K_M",
"ratio": 0.2469582482340622,
"ci": 0.0542806311116404,
"lo": 0.19267761712242182,
"hi": 0.3012388793457026,
"verdict": "LOSS",
"best": 0.32700705186277557
}
},
"git_sha": "1ac219c43d04a69ef4f224e037bef309a4557728",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784681946
},
{
"id": "kernel-perf:llama.cpp/vulkan/evo/Qwen3-1.7B-Q4_K_M:prefill-500",
"kind": "kernel-perf",
"subject": "llama.cpp/vulkan/evo/Qwen3-1.7B-Q4_K_M",
"task": "prefill-500",
"metric": "tok/s",
"value": 4449.614285714286,
"n": 7,
"meta": {
"run": "evo-vulkan-qwen3-1p7b-20260721-175906",
"host": "evo",
"backend": "Vulkan",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 33.58255706047049,
"cv_pct": 0.8160306031067692,
"std": 36.31021429163925,
"samples": [
4478.79,
4377.54,
4439.8,
4447.03,
4446.26,
4482.84,
4475.04
],
"noisy": true
},
"git_sha": "0821c5fcfd729af70037bc1e9e60769d42c081ba",
"git_dirty": false,
"lib_versions": {},
"ts": 1784681946
},
{
"id": "kernel-perf:hanzo-engine/vulkan/evo/Qwen3-1.7B-Q4_K_M:prefill-512",
"kind": "kernel-perf",
"subject": "hanzo-engine/vulkan/evo/Qwen3-1.7B-Q4_K_M",
"task": "prefill-512",
"metric": "tok/s",
"value": 3042.3030767386213,
"n": 6,
"meta": {
"run": "evo-vulkan-qwen3-1p7b-20260721-175906",
"host": "evo",
"backend": "Vulkan",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 10.894437879352038,
"cv_pct": 0.34117397759135737,
"std": 10.3795464172934,
"samples": [
3046.969353951207,
3050.0777734089124,
3046.5896448738617,
3044.1466185631252,
3021.566792976224,
3044.468276658396
],
"noisy": false,
"versus": {
"subject": "llama.cpp/vulkan/evo/Qwen3-1.7B-Q4_K_M",
"ratio": 0.6689791446932548,
"ci": 0.007654633008983679,
"lo": 0.6613245116842711,
"hi": 0.6766337777022385,
"verdict": "LOSS",
"best": 0.6632593633326764
}
},
"git_sha": "1ac219c43d04a69ef4f224e037bef309a4557728",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784681946
},
{
"id": "kernel-perf:llama.cpp/vulkan/evo/Qwen3-1.7B-Q4_K_M:prefill-512",
"kind": "kernel-perf",
"subject": "llama.cpp/vulkan/evo/Qwen3-1.7B-Q4_K_M",
"task": "prefill-512",
"metric": "tok/s",
"value": 4547.68,
"n": 7,
"meta": {
"run": "evo-vulkan-qwen3-1p7b-20260721-175906",
"host": "evo",
"backend": "Vulkan",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 49.42176652419974,
"cv_pct": 1.1750150186066983,
"std": 53.4359229981731,
"samples": [
4598.62,
4575.02,
4583.38,
4555.04,
4546.36,
4538.94,
4436.4
],
"noisy": false
},
"git_sha": "0821c5fcfd729af70037bc1e9e60769d42c081ba",
"git_dirty": false,
"lib_versions": {},
"ts": 1784681946
},
{
"id": "kernel-perf:hanzo-engine/vulkan/evo/Qwen3-1.7B-Q4_K_M:prefill-2048",
"kind": "kernel-perf",
"subject": "hanzo-engine/vulkan/evo/Qwen3-1.7B-Q4_K_M",
"task": "prefill-2048",
"metric": "tok/s",
"value": 1474.989378732038,
"n": 6,
"meta": {
"run": "evo-vulkan-qwen3-1p7b-20260721-175906",
"host": "evo",
"backend": "Vulkan",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 37.626772878907644,
"cv_pct": 2.4304216628223134,
"std": 35.84846138503171,
"samples": [
1402.5038623418957,
1495.9538567890531,
1492.4867662335294,
1489.5727200292106,
1488.1812954191632,
1481.2377715793764
],
"noisy": false,
"versus": {
"subject": "llama.cpp/vulkan/evo/Qwen3-1.7B-Q4_K_M",
"ratio": 0.3465041307658707,
"ci": 0.009146583480855262,
"lo": 0.33735754728501544,
"hi": 0.355650714246726,
"verdict": "LOSS",
"best": 0.3486811838718448
}
},
"git_sha": "1ac219c43d04a69ef4f224e037bef309a4557728",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784681946
},
{
"id": "kernel-perf:llama.cpp/vulkan/evo/Qwen3-1.7B-Q4_K_M:prefill-2048",
"kind": "kernel-perf",
"subject": "llama.cpp/vulkan/evo/Qwen3-1.7B-Q4_K_M",
"task": "prefill-2048",
"metric": "tok/s",
"value": 4256.772857142858,
"n": 7,
"meta": {
"run": "evo-vulkan-qwen3-1p7b-20260721-175906",
"host": "evo",
"backend": "Vulkan",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 28.88198533898117,
"cv_pct": 0.7336038758186615,
"std": 31.22785066479678,
"samples": [
4232.1,
4202.75,
4272.94,
4257.52,
4254.49,
4290.32,
4287.29
],
"noisy": false
},
"git_sha": "0821c5fcfd729af70037bc1e9e60769d42c081ba",
"git_dirty": false,
"lib_versions": {},
"ts": 1784681946
},
{
"id": "kernel-perf:hanzo-engine/vulkan/evo/Qwen3-1.7B-Q4_K_M:prefill-4096",
"kind": "kernel-perf",
"subject": "hanzo-engine/vulkan/evo/Qwen3-1.7B-Q4_K_M",
"task": "prefill-4096",
"metric": "tok/s",
"value": 922.847988818516,
"n": 6,
"meta": {
"run": "evo-vulkan-qwen3-1p7b-20260721-175906",
"host": "evo",
"backend": "Vulkan",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 33.292771524386325,
"cv_pct": 3.437109188689541,
"std": 31.71929302131784,
"samples": [
860.3561548108506,
919.7742104585028,
934.4585318250671,
938.1715332515035,
942.360265632659,
941.9672369325127
],
"noisy": false,
"versus": {
"subject": "llama.cpp/vulkan/evo/Qwen3-1.7B-Q4_K_M",
"ratio": 0.23968156621694192,
"ci": 0.010103308651944356,
"lo": 0.22957825756499756,
"hi": 0.2497848748688863,
"verdict": "LOSS",
"best": 0.23701452369557516
}
},
"git_sha": "1ac219c43d04a69ef4f224e037bef309a4557728",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784681946
},
{
"id": "kernel-perf:llama.cpp/vulkan/evo/Qwen3-1.7B-Q4_K_M:prefill-4096",
"kind": "kernel-perf",
"subject": "llama.cpp/vulkan/evo/Qwen3-1.7B-Q4_K_M",
"task": "prefill-4096",
"metric": "tok/s",
"value": 3850.3085714285708,
"n": 7,
"meta": {
"run": "evo-vulkan-qwen3-1p7b-20260721-175906",
"host": "evo",
"backend": "Vulkan",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 83.95026427109588,
"cv_pct": 2.357445069742959,
"std": 90.76890958703342,
"samples": [
3795.2,
3739.03,
3744.96,
3975.96,
3902.05,
3897.6,
3897.36
],
"noisy": false
},
"git_sha": "0821c5fcfd729af70037bc1e9e60769d42c081ba",
"git_dirty": false,
"lib_versions": {},
"ts": 1784681946
},
{
"id": "kernel-perf:hanzo-engine/vulkan/evo/Qwen3-1.7B-Q4_K_M:decode-128",
"kind": "kernel-perf",
"subject": "hanzo-engine/vulkan/evo/Qwen3-1.7B-Q4_K_M",
"task": "decode-128",
"metric": "tok/s",
"value": 93.90672850690005,
"n": 6,
"meta": {
"run": "evo-vulkan-qwen3-1p7b-20260721-175906",
"host": "evo",
"backend": "Vulkan",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 1.698719495997397,
"cv_pct": 1.7234492870922595,
"std": 1.6184348429838327,
"samples": [
95.61594550593395,
95.29591295143658,
93.29954164006882,
92.22606766839793,
95.03826735665936,
91.96463591890374
],
"noisy": false,
"versus": {
"subject": "llama.cpp/vulkan/evo/Qwen3-1.7B-Q4_K_M",
"ratio": 0.6041693239894012,
"ci": 0.010934153776872084,
"lo": 0.593235170212529,
"hi": 0.6151034777662733,
"verdict": "LOSS",
"best": 0.6144865170076024
}
},
"git_sha": "1ac219c43d04a69ef4f224e037bef309a4557728",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784681946
},
{
"id": "kernel-perf:llama.cpp/vulkan/evo/Qwen3-1.7B-Q4_K_M:decode-128",
"kind": "kernel-perf",
"subject": "llama.cpp/vulkan/evo/Qwen3-1.7B-Q4_K_M",
"task": "decode-128",
"metric": "tok/s",
"value": 155.43114285714287,
"n": 7,
"meta": {
"run": "evo-vulkan-qwen3-1p7b-20260721-175906",
"host": "evo",
"backend": "Vulkan",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "Marketing Name: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 0.08567899917962626,
"cv_pct": 0.05960070476247053,
"std": 0.09263805656321952,
"samples": [
155.322,
155.399,
155.46,
155.344,
155.435,
155.455,
155.603
],
"noisy": false
},
"git_sha": "0821c5fcfd729af70037bc1e9e60769d42c081ba",
"git_dirty": false,
"lib_versions": {},
"ts": 1784681946
},
{
"id": "kernel-perf:hanzo-engine/cuda/spark/Qwen3-1.7B-Q4_K_M:prefill-500",
"kind": "kernel-perf",
"subject": "hanzo-engine/cuda/spark/Qwen3-1.7B-Q4_K_M",
"task": "prefill-500",
"metric": "tok/s",
"value": 7961.3052839552765,
"n": 6,
"meta": {
"run": "spark-cuda-qwen3-1p7b-20260721-182202",
"host": "spark",
"backend": "CUDA",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "NVIDIA GB10, 580.159.03",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 228.6679217365138,
"cv_pct": 2.7364940963124234,
"std": 217.86064908484516,
"samples": [
7768.875337208035,
8078.893432027681,
7794.115386765361,
8083.921713491329,
8282.109452615012,
7759.916381624244
],
"noisy": false,
"versus": {
"subject": "llama.cpp/cuda/spark/Qwen3-1.7B-Q4_K_M",
"ratio": 1.0980740745028479,
"ci": 0.04463029601472079,
"lo": 1.0534437784881272,
"hi": 1.1427043705175686,
"verdict": "WIN",
"best": 1.104438826954647
}
},
"git_sha": "a0c6d7429233bca12e0f4996bbcc9f1c8ca42820",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784683322
},
{
"id": "kernel-perf:llama.cpp/cuda/spark/Qwen3-1.7B-Q4_K_M:prefill-500",
"kind": "kernel-perf",
"subject": "llama.cpp/cuda/spark/Qwen3-1.7B-Q4_K_M",
"task": "prefill-500",
"metric": "tok/s",
"value": 7250.244285714286,
"n": 7,
"meta": {
"run": "spark-cuda-qwen3-1p7b-20260721-182202",
"host": "spark",
"backend": "CUDA",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "NVIDIA GB10, 580.159.03",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 208.49594824668517,
"cv_pct": 3.1092814123566384,
"std": 225.4304979261636,
"samples": [
6839.31,
7177.51,
7369.08,
7300.43,
7124.96,
7441.49,
7498.93
],
"noisy": false
},
"git_sha": "31e82494c0a3913c919c1027fa70500fbf4c07dd",
"git_dirty": false,
"lib_versions": {},
"ts": 1784683322
},
{
"id": "kernel-perf:hanzo-engine/cuda/spark/Qwen3-1.7B-Q4_K_M:prefill-512",
"kind": "kernel-perf",
"subject": "hanzo-engine/cuda/spark/Qwen3-1.7B-Q4_K_M",
"task": "prefill-512",
"metric": "tok/s",
"value": 7873.617764504105,
"n": 6,
"meta": {
"run": "spark-cuda-qwen3-1p7b-20260721-182202",
"host": "spark",
"backend": "CUDA",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "NVIDIA GB10, 580.159.03",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 608.171872746476,
"cv_pct": 7.359114356142873,
"std": 579.4285352554372,
"samples": [
8866.333152849193,
7363.158891806045,
7322.247298376772,
7659.644532951902,
8161.40075396359,
7868.92195707712
],
"noisy": true,
"versus": {
"subject": "llama.cpp/cuda/spark/Qwen3-1.7B-Q4_K_M",
"ratio": 1.2719858359915002,
"ci": 0.24302958316006026,
"lo": 1.02895625283144,
"hi": 1.5150154191515603,
"verdict": "WIN",
"best": 1.2038862302164894
}
},
"git_sha": "a0c6d7429233bca12e0f4996bbcc9f1c8ca42820",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784683322
},
{
"id": "kernel-perf:llama.cpp/cuda/spark/Qwen3-1.7B-Q4_K_M:prefill-512",
"kind": "kernel-perf",
"subject": "llama.cpp/cuda/spark/Qwen3-1.7B-Q4_K_M",
"task": "prefill-512",
"metric": "tok/s",
"value": 6190.0199999999995,
"n": 7,
"meta": {
"run": "spark-cuda-qwen3-1p7b-20260721-182202",
"host": "spark",
"backend": "CUDA",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "NVIDIA GB10, 580.159.03",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 1081.728459378465,
"cv_pct": 18.89475453127192,
"std": 1169.5890844366381,
"samples": [
6120.88,
4768.52,
5096.15,
5259.2,
7364.76,
7363.22,
7357.41
],
"noisy": true
},
"git_sha": "31e82494c0a3913c919c1027fa70500fbf4c07dd",
"git_dirty": false,
"lib_versions": {},
"ts": 1784683322
},
{
"id": "kernel-perf:hanzo-engine/cuda/spark/Qwen3-1.7B-Q4_K_M:prefill-2048",
"kind": "kernel-perf",
"subject": "hanzo-engine/cuda/spark/Qwen3-1.7B-Q4_K_M",
"task": "prefill-2048",
"metric": "tok/s",
"value": 7285.553351380399,
"n": 6,
"meta": {
"run": "spark-cuda-qwen3-1p7b-20260721-182202",
"host": "spark",
"backend": "CUDA",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "NVIDIA GB10, 580.159.03",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 745.1504911314897,
"cv_pct": 9.744397636087628,
"std": 709.9332885478146,
"samples": [
7933.1939232028935,
7516.555828970682,
7542.055105060772,
7846.8565182438615,
6774.52638882455,
6100.132343979636
],
"noisy": true,
"versus": {
"subject": "llama.cpp/cuda/spark/Qwen3-1.7B-Q4_K_M",
"ratio": 0.9943646251779228,
"ci": 0.11489964542135103,
"lo": 0.8794649797565718,
"hi": 1.109264270599274,
"verdict": "PARITY",
"best": 0.9861048657924437
}
},
"git_sha": "a0c6d7429233bca12e0f4996bbcc9f1c8ca42820",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784683322
},
{
"id": "kernel-perf:llama.cpp/cuda/spark/Qwen3-1.7B-Q4_K_M:prefill-2048",
"kind": "kernel-perf",
"subject": "llama.cpp/cuda/spark/Qwen3-1.7B-Q4_K_M",
"task": "prefill-2048",
"metric": "tok/s",
"value": 7326.8428571428585,
"n": 7,
"meta": {
"run": "spark-cuda-qwen3-1p7b-20260721-182202",
"host": "spark",
"backend": "CUDA",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "NVIDIA GB10, 580.159.03",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 393.9661616687155,
"cv_pct": 5.813759936771857,
"std": 425.965054658802,
"samples": [
6893.29,
7308.59,
7134.73,
8044.98,
7103.14,
7023.08,
7780.09
],
"noisy": true
},
"git_sha": "31e82494c0a3913c919c1027fa70500fbf4c07dd",
"git_dirty": false,
"lib_versions": {},
"ts": 1784683322
},
{
"id": "kernel-perf:hanzo-engine/cuda/spark/Qwen3-1.7B-Q4_K_M:prefill-4096",
"kind": "kernel-perf",
"subject": "hanzo-engine/cuda/spark/Qwen3-1.7B-Q4_K_M",
"task": "prefill-4096",
"metric": "tok/s",
"value": 7474.835466369842,
"n": 6,
"meta": {
"run": "spark-cuda-qwen3-1p7b-20260721-182202",
"host": "spark",
"backend": "CUDA",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "NVIDIA GB10, 580.159.03",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 217.09069019906974,
"cv_pct": 2.767025174564947,
"std": 206.8305791117627,
"samples": [
7236.708918351875,
7734.218783054474,
7356.942316830457,
7481.7421745178035,
7709.355621697786,
7330.044983766661
],
"noisy": false,
"versus": {
"subject": "llama.cpp/cuda/spark/Qwen3-1.7B-Q4_K_M",
"ratio": 1.0878403542162853,
"ci": 0.03727819406280123,
"lo": 1.050562160153484,
"hi": 1.1251185482790866,
"verdict": "WIN",
"best": 1.094126878212789
}
},
"git_sha": "a0c6d7429233bca12e0f4996bbcc9f1c8ca42820",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784683322
},
{
"id": "kernel-perf:llama.cpp/cuda/spark/Qwen3-1.7B-Q4_K_M:prefill-4096",
"kind": "kernel-perf",
"subject": "llama.cpp/cuda/spark/Qwen3-1.7B-Q4_K_M",
"task": "prefill-4096",
"metric": "tok/s",
"value": 6871.261428571429,
"n": 7,
"meta": {
"run": "spark-cuda-qwen3-1p7b-20260721-182202",
"host": "spark",
"backend": "CUDA",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "NVIDIA GB10, 580.159.03",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 124.97623176004703,
"cv_pct": 1.9665545790962498,
"std": 135.12710626524583,
"samples": [
7017.83,
6727.14,
6815.02,
7068.85,
6811.36,
6924.16,
6734.47
],
"noisy": false
},
"git_sha": "31e82494c0a3913c919c1027fa70500fbf4c07dd",
"git_dirty": false,
"lib_versions": {},
"ts": 1784683322
},
{
"id": "kernel-perf:hanzo-engine/cuda/spark/Qwen3-1.7B-Q4_K_M:decode-128",
"kind": "kernel-perf",
"subject": "hanzo-engine/cuda/spark/Qwen3-1.7B-Q4_K_M",
"task": "decode-128",
"metric": "tok/s",
"value": 86.19569259011509,
"n": 6,
"meta": {
"run": "spark-cuda-qwen3-1p7b-20260721-182202",
"host": "spark",
"backend": "CUDA",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "NVIDIA GB10, 580.159.03",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 0.9595707936609261,
"cv_pct": 1.0606326836321716,
"std": 0.9142196874938745,
"samples": [
85.77060016258,
86.92451958895315,
87.15579114755651,
85.03349629442704,
85.37369283103676,
86.91605551613715
],
"noisy": false,
"versus": {
"subject": "llama.cpp/cuda/spark/Qwen3-1.7B-Q4_K_M",
"ratio": 0.8385331026313599,
"ci": 0.013118971126756926,
"lo": 0.825414131504603,
"hi": 0.8516520737581169,
"verdict": "LOSS",
"best": 0.8480551540029299
}
},
"git_sha": "a0c6d7429233bca12e0f4996bbcc9f1c8ca42820",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784683322
},
{
"id": "kernel-perf:llama.cpp/cuda/spark/Qwen3-1.7B-Q4_K_M:decode-128",
"kind": "kernel-perf",
"subject": "llama.cpp/cuda/spark/Qwen3-1.7B-Q4_K_M",
"task": "decode-128",
"metric": "tok/s",
"value": 102.79342857142858,
"n": 7,
"meta": {
"run": "spark-cuda-qwen3-1p7b-20260721-182202",
"host": "spark",
"backend": "CUDA",
"model": "Qwen3-1.7B-Q4_K_M",
"model_sha256": "72c5c3cb38fa32d5256e2fe30d03e7a64c6c79e668ad84057e3bd66e250b24fb",
"model_bytes": 1282439584,
"gpu": "NVIDIA GB10, 580.159.03",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 1.1299737322435264,
"cv_pct": 1.1885516156715044,
"std": 1.2217529560898481,
"samples": [
100.938,
103.878,
102.381,
102.05,
102.233,
103.715,
104.359
],
"noisy": false
},
"git_sha": "31e82494c0a3913c919c1027fa70500fbf4c07dd",
"git_dirty": false,
"lib_versions": {},
"ts": 1784683322
},
{
"id": "kernel-perf:hanzo-engine/cuda/spark/Qwen3-4B-Q4_K_M:prefill-512",
"kind": "kernel-perf",
"subject": "hanzo-engine/cuda/spark/Qwen3-4B-Q4_K_M",
"task": "prefill-512",
"metric": "tok/s",
"value": 3857.5678042485233,
"n": 6,
"meta": {
"run": "spark-cuda-qwen3-4b-20260721-183326",
"host": "spark",
"backend": "CUDA",
"model": "Qwen3-4B-Q4_K_M",
"model_sha256": "7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5",
"model_bytes": 2497280256,
"gpu": "NVIDIA GB10, 580.159.03",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 336.44384062121463,
"cv_pct": 8.309455418811957,
"std": 320.54287694447436,
"samples": [
3745.424903129742,
4076.446362978631,
4309.637596261585,
3977.1762593866506,
3565.577552182262,
3471.1441515522692
],
"noisy": true,
"versus": {
"subject": "llama.cpp/cuda/spark/Qwen3-4B-Q4_K_M",
"ratio": 1.0574034219185633,
"ci": 0.11428596616108816,
"lo": 0.9431174557574751,
"hi": 1.1716893880796515,
"verdict": "PARITY",
"best": 1.0840955584219636
}
},
"git_sha": "a0c6d7429233bca12e0f4996bbcc9f1c8ca42820",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784684006
},
{
"id": "kernel-perf:llama.cpp/cuda/spark/Qwen3-4B-Q4_K_M:prefill-512",
"kind": "kernel-perf",
"subject": "llama.cpp/cuda/spark/Qwen3-4B-Q4_K_M",
"task": "prefill-512",
"metric": "tok/s",
"value": 3648.1514285714284,
"n": 7,
"meta": {
"run": "spark-cuda-qwen3-4b-20260721-183326",
"host": "spark",
"backend": "CUDA",
"model": "Qwen3-4B-Q4_K_M",
"model_sha256": "7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5",
"model_bytes": 2497280256,
"gpu": "NVIDIA GB10, 580.159.03",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 232.88024860451574,
"cv_pct": 6.901998329580921,
"std": 251.79535066058247,
"samples": [
3261.61,
3889.38,
3975.33,
3635.09,
3470.81,
3521.85,
3782.99
],
"noisy": true
},
"git_sha": "31e82494c0a3913c919c1027fa70500fbf4c07dd",
"git_dirty": false,
"lib_versions": {},
"ts": 1784684006
},
{
"id": "kernel-perf:hanzo-engine/cuda/spark/Qwen3-4B-Q4_K_M:prefill-2048",
"kind": "kernel-perf",
"subject": "hanzo-engine/cuda/spark/Qwen3-4B-Q4_K_M",
"task": "prefill-2048",
"metric": "tok/s",
"value": 3774.374154151353,
"n": 6,
"meta": {
"run": "spark-cuda-qwen3-4b-20260721-183326",
"host": "spark",
"backend": "CUDA",
"model": "Qwen3-4B-Q4_K_M",
"model_sha256": "7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5",
"model_bytes": 2497280256,
"gpu": "NVIDIA GB10, 580.159.03",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 89.04689761884853,
"cv_pct": 2.2477468215990695,
"std": 84.8383750851938,
"samples": [
3807.6484151911,
3826.504807207347,
3885.08917703372,
3774.6032737602336,
3693.928461137649,
3658.4707905780674
],
"noisy": false,
"versus": {
"subject": "llama.cpp/cuda/spark/Qwen3-4B-Q4_K_M",
"ratio": 1.0506127003875274,
"ci": 0.033923633539275466,
"lo": 1.016689066848252,
"hi": 1.0845363339268028,
"verdict": "WIN",
"best": 1.0421320639464702
}
},
"git_sha": "a0c6d7429233bca12e0f4996bbcc9f1c8ca42820",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784684006
},
{
"id": "kernel-perf:llama.cpp/cuda/spark/Qwen3-4B-Q4_K_M:prefill-2048",
"kind": "kernel-perf",
"subject": "llama.cpp/cuda/spark/Qwen3-4B-Q4_K_M",
"task": "prefill-2048",
"metric": "tok/s",
"value": 3592.5457142857144,
"n": 7,
"meta": {
"run": "spark-cuda-qwen3-4b-20260721-183326",
"host": "spark",
"backend": "CUDA",
"model": "Qwen3-4B-Q4_K_M",
"model_sha256": "7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5",
"model_bytes": 2497280256,
"gpu": "NVIDIA GB10, 580.159.03",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 79.199001503209,
"cv_pct": 2.3835949344443224,
"std": 85.63173766331089,
"samples": [
3601.57,
3533.7,
3498.13,
3663.48,
3728.02,
3505.69,
3617.23
],
"noisy": false
},
"git_sha": "31e82494c0a3913c919c1027fa70500fbf4c07dd",
"git_dirty": false,
"lib_versions": {},
"ts": 1784684006
},
{
"id": "kernel-perf:hanzo-engine/cuda/spark/Qwen3-4B-Q4_K_M:decode-128",
"kind": "kernel-perf",
"subject": "hanzo-engine/cuda/spark/Qwen3-4B-Q4_K_M",
"task": "decode-128",
"metric": "tok/s",
"value": 47.285587260908976,
"n": 6,
"meta": {
"run": "spark-cuda-qwen3-4b-20260721-183326",
"host": "spark",
"backend": "CUDA",
"model": "Qwen3-4B-Q4_K_M",
"model_sha256": "7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5",
"model_bytes": 2497280256,
"gpu": "NVIDIA GB10, 580.159.03",
"sampler": null,
"reps": 7,
"concurrency": 1,
"args": "",
"ci95": 3.858303795297479,
"cv_pct": 7.773939950154127,
"std": 3.6759531587407936,
"samples": [
42.1817514939486,
45.34289482407785,
44.6941243891354,
50.35532246012468,
50.447203876395875,
50.692226521771445
],
"noisy": true,
"versus": {
"subject": "llama.cpp/cuda/spark/Qwen3-4B-Q4_K_M",
"ratio": 0.8803341098371691,
"ci": 0.07333904683182807,
"lo": 0.8069950630053411,
"hi": 0.9536731566689971,
"verdict": "LOSS",
"best": 0.9141492423619224
}
},
"git_sha": "a0c6d7429233bca12e0f4996bbcc9f1c8ca42820",
"git_dirty": false,
"lib_versions": {
"hanzo-engine": "1.7.87",
"hanzo-metal-kernels": "0.11.44",
"hanzo-ml": "0.11.88",
"hanzo-rocm-kernels": "0.11.36"
},
"ts": 1784684006
},
{
"id": "kernel-perf:llama.cpp/cuda/spark/Qwen3-4B-Q4_K_M:decode-128",
"kind": "kernel-perf",
"subject": "llama.cpp/cuda/spark/Qwen3-4B-Q4_K_M",
"task": "decode-128",
"metric": "tok/s",
"value": 53.71322857142858,
"n": 7,
"meta": {
"run": "spark-cuda-qwen3-4b-20260721-183326",
"host": "spark",
"backend": "CUDA",
"model": "Qwen3-4B-Q4_K_M",
"model_sha256": "7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5",
"model_bytes": 2497280256,
"gpu": "NVIDIA GB10, 580.159.03",
"sampler": null,
"reps": 7,
"concurrency": 1,
"ci95": 0.9026160200079303,
"cv_pct": 1.81692428780393,
"std": 0.9759286956779257,
"samples": [
53.094,
54.661,
52.8834,
53.5043,
55.4529,
53.4671,
52.9299
],
"noisy": true
},
"git_sha": "31e82494c0a3913c919c1027fa70500fbf4c07dd",
"git_dirty": false,
"lib_versions": {},
"ts": 1784684006
}
],
"attempts": []
}