Add AIST-87M throughput benchmark
Browse files- README.md +17 -4
- aist87m_vs_dual_audio_throughput_l4_20260504.json +320 -0
- manifest.json +20 -3
README.md
CHANGED
|
@@ -136,10 +136,22 @@ Exact-gate tradeoff against the same dual-audio local baseline:
|
|
| 136 |
| SALT audio-text R@1 avg | 0.008 | 0.007 | flat |
|
| 137 |
| SALT image-audio R@1 avg | 0.138 | 0.148 | -0.010 |
|
| 138 |
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 143 |
|
| 144 |
## Architecture
|
| 145 |
|
|
@@ -175,6 +187,7 @@ The audio encoder in this artifact is the merged native checkpoint:
|
|
| 175 |
| `parameter_breakdown.json` | Exact parameter accounting |
|
| 176 |
| `aist87m_memory_slice_release_report.md` | Human-memory slice report |
|
| 177 |
| `aist87m_memory_slice_release_report.json` | Machine-readable evaluation summary |
|
|
|
|
| 178 |
|
| 179 |
## Caveats
|
| 180 |
|
|
|
|
| 136 |
| SALT audio-text R@1 avg | 0.008 | 0.007 | flat |
|
| 137 |
| SALT image-audio R@1 avg | 0.138 | 0.148 | -0.010 |
|
| 138 |
|
| 139 |
+
Measured PyTorch audio-stack throughput on an NVIDIA L4, using synthetic 10s
|
| 140 |
+
32 kHz CPU waveforms passed through waveform -> audio encoder -> projection ->
|
| 141 |
+
normalized embedding. Median wall time is over 50 timed iterations after 20
|
| 142 |
+
warmup iterations. This excludes audio file decode, dataset download, and MTEB
|
| 143 |
+
result serialization.
|
| 144 |
+
|
| 145 |
+
| Batch | AIST-87M median ms | AIST-87M throughput | AIST-95M median ms | AIST-95M throughput | Speedup |
|
| 146 |
+
|---:|---:|---:|---:|---:|---:|
|
| 147 |
+
| 1 | 5.36 | 186.7 clips/s; 1,867 audio-s/s | 10.50 | 95.2 clips/s; 952 audio-s/s | 1.96x |
|
| 148 |
+
| 8 | 16.46 | 486.0 clips/s; 4,860 audio-s/s | 60.29 | 132.7 clips/s; 1,327 audio-s/s | 3.66x |
|
| 149 |
+
| 16 | 41.19 | 388.5 clips/s; 3,885 audio-s/s | 133.95 | 119.4 clips/s; 1,194 audio-s/s | 3.25x |
|
| 150 |
+
|
| 151 |
+
Projection-only throughput at feature batch 2048 is also higher for the
|
| 152 |
+
single-audio path: 314k features/s for `AIST-87M` vs 282k features/s for the
|
| 153 |
+
dual-audio tower. Raw benchmark output is included as
|
| 154 |
+
`aist87m_vs_dual_audio_throughput_l4_20260504.json`.
|
| 155 |
|
| 156 |
## Architecture
|
| 157 |
|
|
|
|
| 187 |
| `parameter_breakdown.json` | Exact parameter accounting |
|
| 188 |
| `aist87m_memory_slice_release_report.md` | Human-memory slice report |
|
| 189 |
| `aist87m_memory_slice_release_report.json` | Machine-readable evaluation summary |
|
| 190 |
+
| `aist87m_vs_dual_audio_throughput_l4_20260504.json` | L4 throughput benchmark vs dual-audio tower |
|
| 191 |
|
| 192 |
## Caveats
|
| 193 |
|
aist87m_vs_dual_audio_throughput_l4_20260504.json
ADDED
|
@@ -0,0 +1,320 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"aist87": {
|
| 3 |
+
"audio_adapter_params": 0,
|
| 4 |
+
"audio_checkpoint": "triembed/checkpoints/mn20_native_merged_aistmix_audioheavy100k175k175k_continue_from_balanced_20260426T143137Z/latest_model.pt",
|
| 5 |
+
"audio_head_input_dim": 1280,
|
| 6 |
+
"audio_path_params": 32144280,
|
| 7 |
+
"audio_projection_params": 12306560,
|
| 8 |
+
"audio_sample_rate": 32000,
|
| 9 |
+
"audio_seconds": 10.0,
|
| 10 |
+
"checkpoint": "triembed/checkpoints/aist81m_raw1280_mn20_lora_teacher_20260502T001744Z/best_model.pt",
|
| 11 |
+
"config_run_name": "aist81m_raw1280_mn20_lora_teacher_20260502T001744Z",
|
| 12 |
+
"encoder_params": 19837720,
|
| 13 |
+
"name": "AIST-87M",
|
| 14 |
+
"projection_only": {
|
| 15 |
+
"batch_size": 2048,
|
| 16 |
+
"cuda_events": {
|
| 17 |
+
"max_ms": 6.870016098022461,
|
| 18 |
+
"mean_ms": 6.54273796081543,
|
| 19 |
+
"median_ms": 6.502351999282837,
|
| 20 |
+
"min_ms": 6.4389119148254395,
|
| 21 |
+
"p95_ms": 6.860799789428711
|
| 22 |
+
},
|
| 23 |
+
"features_per_sec_cuda": 313018.8022606908,
|
| 24 |
+
"features_per_sec_cuda_median": 314962.9549778111,
|
| 25 |
+
"features_per_sec_wall": 312183.5609238512,
|
| 26 |
+
"features_per_sec_wall_median": 314158.8298674527,
|
| 27 |
+
"wall": {
|
| 28 |
+
"max_ms": 6.898713065311313,
|
| 29 |
+
"mean_ms": 6.5602429350838065,
|
| 30 |
+
"median_ms": 6.51899550575763,
|
| 31 |
+
"min_ms": 6.456336006522179,
|
| 32 |
+
"p95_ms": 6.8806440103799105
|
| 33 |
+
}
|
| 34 |
+
},
|
| 35 |
+
"waveform_to_embedding": {
|
| 36 |
+
"1": {
|
| 37 |
+
"audio_seconds_per_sec_cuda": 1767.600123842052,
|
| 38 |
+
"audio_seconds_per_sec_cuda_median": 1872.479629331263,
|
| 39 |
+
"audio_seconds_per_sec_wall": 1762.1861633836697,
|
| 40 |
+
"audio_seconds_per_sec_wall_median": 1866.9151048991507,
|
| 41 |
+
"cuda_events": {
|
| 42 |
+
"max_ms": 6.4194560050964355,
|
| 43 |
+
"mean_ms": 5.657388153076172,
|
| 44 |
+
"median_ms": 5.340512037277222,
|
| 45 |
+
"min_ms": 5.30841588973999,
|
| 46 |
+
"p95_ms": 6.403071880340576
|
| 47 |
+
},
|
| 48 |
+
"items_per_sec_cuda": 176.7600123842052,
|
| 49 |
+
"items_per_sec_cuda_median": 187.2479629331263,
|
| 50 |
+
"items_per_sec_wall": 176.21861633836699,
|
| 51 |
+
"items_per_sec_wall_median": 186.69151048991506,
|
| 52 |
+
"wall": {
|
| 53 |
+
"max_ms": 6.437776144593954,
|
| 54 |
+
"mean_ms": 5.674769333563745,
|
| 55 |
+
"median_ms": 5.3564299596473575,
|
| 56 |
+
"min_ms": 5.324810044839978,
|
| 57 |
+
"p95_ms": 6.428366061300039
|
| 58 |
+
}
|
| 59 |
+
},
|
| 60 |
+
"16": {
|
| 61 |
+
"audio_seconds_per_sec_cuda": 3615.280314396054,
|
| 62 |
+
"audio_seconds_per_sec_cuda_median": 3886.776599612583,
|
| 63 |
+
"audio_seconds_per_sec_wall": 3613.4877550602714,
|
| 64 |
+
"audio_seconds_per_sec_wall_median": 3884.636076443369,
|
| 65 |
+
"cuda_events": {
|
| 66 |
+
"max_ms": 92.48665618896484,
|
| 67 |
+
"mean_ms": 44.25659591674805,
|
| 68 |
+
"median_ms": 41.16521644592285,
|
| 69 |
+
"min_ms": 40.92313766479492,
|
| 70 |
+
"p95_ms": 68.5506591796875
|
| 71 |
+
},
|
| 72 |
+
"items_per_sec_cuda": 361.5280314396054,
|
| 73 |
+
"items_per_sec_cuda_median": 388.6776599612583,
|
| 74 |
+
"items_per_sec_wall": 361.34877550602715,
|
| 75 |
+
"items_per_sec_wall_median": 388.4636076443369,
|
| 76 |
+
"wall": {
|
| 77 |
+
"max_ms": 92.50957006588578,
|
| 78 |
+
"mean_ms": 44.278550487942994,
|
| 79 |
+
"median_ms": 41.187899420037866,
|
| 80 |
+
"min_ms": 40.94505496323109,
|
| 81 |
+
"p95_ms": 68.57325998134911
|
| 82 |
+
}
|
| 83 |
+
},
|
| 84 |
+
"8": {
|
| 85 |
+
"audio_seconds_per_sec_cuda": 4831.796135966214,
|
| 86 |
+
"audio_seconds_per_sec_cuda_median": 4866.393306722875,
|
| 87 |
+
"audio_seconds_per_sec_wall": 4825.5348259899165,
|
| 88 |
+
"audio_seconds_per_sec_wall_median": 4859.857786409293,
|
| 89 |
+
"cuda_events": {
|
| 90 |
+
"max_ms": 18.59174346923828,
|
| 91 |
+
"mean_ms": 16.556989936828614,
|
| 92 |
+
"median_ms": 16.439279556274414,
|
| 93 |
+
"min_ms": 16.22015953063965,
|
| 94 |
+
"p95_ms": 16.830463409423828
|
| 95 |
+
},
|
| 96 |
+
"items_per_sec_cuda": 483.17961359662144,
|
| 97 |
+
"items_per_sec_cuda_median": 486.6393306722875,
|
| 98 |
+
"items_per_sec_wall": 482.5534825989917,
|
| 99 |
+
"items_per_sec_wall_median": 485.9857786409293,
|
| 100 |
+
"wall": {
|
| 101 |
+
"max_ms": 18.62559001892805,
|
| 102 |
+
"mean_ms": 16.57847324386239,
|
| 103 |
+
"median_ms": 16.46138704381883,
|
| 104 |
+
"min_ms": 16.24572789296508,
|
| 105 |
+
"p95_ms": 16.852295957505703
|
| 106 |
+
}
|
| 107 |
+
}
|
| 108 |
+
}
|
| 109 |
+
},
|
| 110 |
+
"comparison_by_batch_size": {
|
| 111 |
+
"1": {
|
| 112 |
+
"aist87_audio_seconds_per_sec_wall": 1762.1861633836697,
|
| 113 |
+
"aist87_audio_seconds_per_sec_wall_median": 1866.9151048991507,
|
| 114 |
+
"aist87_items_per_sec_cuda": 176.7600123842052,
|
| 115 |
+
"aist87_items_per_sec_cuda_median": 187.2479629331263,
|
| 116 |
+
"aist87_items_per_sec_wall": 176.21861633836699,
|
| 117 |
+
"aist87_items_per_sec_wall_median": 186.69151048991506,
|
| 118 |
+
"aist87_mean_cuda_ms": 5.657388153076172,
|
| 119 |
+
"aist87_mean_wall_ms": 5.674769333563745,
|
| 120 |
+
"aist87_median_cuda_ms": 5.340512037277222,
|
| 121 |
+
"aist87_median_wall_ms": 5.3564299596473575,
|
| 122 |
+
"dual_audio_seconds_per_sec_wall": 841.2845368070198,
|
| 123 |
+
"dual_audio_seconds_per_sec_wall_median": 952.2081950454711,
|
| 124 |
+
"dual_items_per_sec_cuda": 84.26473443749266,
|
| 125 |
+
"dual_items_per_sec_cuda_median": 95.38999079009756,
|
| 126 |
+
"dual_items_per_sec_wall": 84.12845368070198,
|
| 127 |
+
"dual_items_per_sec_wall_median": 95.22081950454712,
|
| 128 |
+
"dual_mean_cuda_ms": 11.867360725402833,
|
| 129 |
+
"dual_mean_wall_ms": 11.8865848146379,
|
| 130 |
+
"dual_median_cuda_ms": 10.483280181884766,
|
| 131 |
+
"dual_median_wall_ms": 10.501904995180666,
|
| 132 |
+
"speedup_audio_seconds_per_sec_wall": 2.094637529024135,
|
| 133 |
+
"speedup_audio_seconds_per_sec_wall_median": 1.960616506571863,
|
| 134 |
+
"speedup_items_per_sec_cuda": 2.0976748287900353,
|
| 135 |
+
"speedup_items_per_sec_cuda_median": 1.962972858915136,
|
| 136 |
+
"speedup_items_per_sec_wall": 2.094637529024135,
|
| 137 |
+
"speedup_items_per_sec_wall_median": 1.9606165065718626
|
| 138 |
+
},
|
| 139 |
+
"16": {
|
| 140 |
+
"aist87_audio_seconds_per_sec_wall": 3613.4877550602714,
|
| 141 |
+
"aist87_audio_seconds_per_sec_wall_median": 3884.636076443369,
|
| 142 |
+
"aist87_items_per_sec_cuda": 361.5280314396054,
|
| 143 |
+
"aist87_items_per_sec_cuda_median": 388.6776599612583,
|
| 144 |
+
"aist87_items_per_sec_wall": 361.34877550602715,
|
| 145 |
+
"aist87_items_per_sec_wall_median": 388.4636076443369,
|
| 146 |
+
"aist87_mean_cuda_ms": 44.25659591674805,
|
| 147 |
+
"aist87_mean_wall_ms": 44.278550487942994,
|
| 148 |
+
"aist87_median_cuda_ms": 41.16521644592285,
|
| 149 |
+
"aist87_median_wall_ms": 41.187899420037866,
|
| 150 |
+
"dual_audio_seconds_per_sec_wall": 1173.9429007331944,
|
| 151 |
+
"dual_audio_seconds_per_sec_wall_median": 1194.4542069368163,
|
| 152 |
+
"dual_items_per_sec_cuda": 117.41469647216593,
|
| 153 |
+
"dual_items_per_sec_cuda_median": 119.4681483517097,
|
| 154 |
+
"dual_items_per_sec_wall": 117.39429007331943,
|
| 155 |
+
"dual_items_per_sec_wall_median": 119.44542069368163,
|
| 156 |
+
"dual_mean_cuda_ms": 136.26914245605468,
|
| 157 |
+
"dual_mean_wall_ms": 136.29282983019948,
|
| 158 |
+
"dual_median_cuda_ms": 133.92691040039062,
|
| 159 |
+
"dual_median_wall_ms": 133.95239354576916,
|
| 160 |
+
"speedup_audio_seconds_per_sec_wall": 3.078077948087119,
|
| 161 |
+
"speedup_audio_seconds_per_sec_wall_median": 3.252226878086467,
|
| 162 |
+
"speedup_items_per_sec_cuda": 3.079069676131288,
|
| 163 |
+
"speedup_items_per_sec_cuda_median": 3.2533998837665585,
|
| 164 |
+
"speedup_items_per_sec_wall": 3.0780779480871194,
|
| 165 |
+
"speedup_items_per_sec_wall_median": 3.252226878086467
|
| 166 |
+
},
|
| 167 |
+
"8": {
|
| 168 |
+
"aist87_audio_seconds_per_sec_wall": 4825.5348259899165,
|
| 169 |
+
"aist87_audio_seconds_per_sec_wall_median": 4859.857786409293,
|
| 170 |
+
"aist87_items_per_sec_cuda": 483.17961359662144,
|
| 171 |
+
"aist87_items_per_sec_cuda_median": 486.6393306722875,
|
| 172 |
+
"aist87_items_per_sec_wall": 482.5534825989917,
|
| 173 |
+
"aist87_items_per_sec_wall_median": 485.9857786409293,
|
| 174 |
+
"aist87_mean_cuda_ms": 16.556989936828614,
|
| 175 |
+
"aist87_mean_wall_ms": 16.57847324386239,
|
| 176 |
+
"aist87_median_cuda_ms": 16.439279556274414,
|
| 177 |
+
"aist87_median_wall_ms": 16.46138704381883,
|
| 178 |
+
"dual_audio_seconds_per_sec_wall": 1304.9761567848066,
|
| 179 |
+
"dual_audio_seconds_per_sec_wall_median": 1326.985212046842,
|
| 180 |
+
"dual_items_per_sec_cuda": 130.54346670818657,
|
| 181 |
+
"dual_items_per_sec_cuda_median": 132.74486741223745,
|
| 182 |
+
"dual_items_per_sec_wall": 130.49761567848066,
|
| 183 |
+
"dual_items_per_sec_wall_median": 132.6985212046842,
|
| 184 |
+
"dual_mean_cuda_ms": 61.28227020263672,
|
| 185 |
+
"dual_mean_wall_ms": 61.30380205344409,
|
| 186 |
+
"dual_median_cuda_ms": 60.26598358154297,
|
| 187 |
+
"dual_median_wall_ms": 60.28703204356134,
|
| 188 |
+
"speedup_audio_seconds_per_sec_wall": 3.6977953971810824,
|
| 189 |
+
"speedup_audio_seconds_per_sec_wall_median": 3.6623300262051015,
|
| 190 |
+
"speedup_items_per_sec_cuda": 3.7012929546042197,
|
| 191 |
+
"speedup_items_per_sec_cuda_median": 3.6659747390536417,
|
| 192 |
+
"speedup_items_per_sec_wall": 3.697795397181083,
|
| 193 |
+
"speedup_items_per_sec_wall_median": 3.6623300262051015
|
| 194 |
+
}
|
| 195 |
+
},
|
| 196 |
+
"dual_audio_tower": {
|
| 197 |
+
"audio_adapter_params": 0,
|
| 198 |
+
"audio_head_input_dim": 2304,
|
| 199 |
+
"audio_path_params": 40390311,
|
| 200 |
+
"audio_projection_params": 14272640,
|
| 201 |
+
"audio_sample_rate": 32000,
|
| 202 |
+
"audio_seconds": 10.0,
|
| 203 |
+
"checkpoint": "triembed/checkpoints/wordnet_validaudio_real_mn20_whisper_wavcaps_speech150k_d2_retrieval_20260418T004617Z/best_model.pt",
|
| 204 |
+
"config_run_name": "wordnet_validaudio_real_mn20_whisper_wavcaps_speech150k_d2_retrieval_20260418T004617Z",
|
| 205 |
+
"encoder_params": 26117671,
|
| 206 |
+
"name": "AIST-95M dual-audio tower",
|
| 207 |
+
"projection_only": {
|
| 208 |
+
"batch_size": 2048,
|
| 209 |
+
"cuda_events": {
|
| 210 |
+
"max_ms": 7.5786237716674805,
|
| 211 |
+
"mean_ms": 7.27663553237915,
|
| 212 |
+
"median_ms": 7.247359991073608,
|
| 213 |
+
"min_ms": 7.027616024017334,
|
| 214 |
+
"p95_ms": 7.532544136047363
|
| 215 |
+
},
|
| 216 |
+
"features_per_sec_cuda": 281448.75346400525,
|
| 217 |
+
"features_per_sec_cuda_median": 282585.6591258707,
|
| 218 |
+
"features_per_sec_wall": 280732.89089042274,
|
| 219 |
+
"features_per_sec_wall_median": 281735.96789576364,
|
| 220 |
+
"wall": {
|
| 221 |
+
"max_ms": 7.597371004521847,
|
| 222 |
+
"mean_ms": 7.295190789736807,
|
| 223 |
+
"median_ms": 7.26921739988029,
|
| 224 |
+
"min_ms": 7.046802900731564,
|
| 225 |
+
"p95_ms": 7.553461939096451
|
| 226 |
+
}
|
| 227 |
+
},
|
| 228 |
+
"waveform_to_embedding": {
|
| 229 |
+
"1": {
|
| 230 |
+
"audio_seconds_per_sec_cuda": 842.6473443749267,
|
| 231 |
+
"audio_seconds_per_sec_cuda_median": 953.8999079009756,
|
| 232 |
+
"audio_seconds_per_sec_wall": 841.2845368070198,
|
| 233 |
+
"audio_seconds_per_sec_wall_median": 952.2081950454711,
|
| 234 |
+
"cuda_events": {
|
| 235 |
+
"max_ms": 53.797889709472656,
|
| 236 |
+
"mean_ms": 11.867360725402833,
|
| 237 |
+
"median_ms": 10.483280181884766,
|
| 238 |
+
"min_ms": 10.369024276733398,
|
| 239 |
+
"p95_ms": 10.95680046081543
|
| 240 |
+
},
|
| 241 |
+
"items_per_sec_cuda": 84.26473443749266,
|
| 242 |
+
"items_per_sec_cuda_median": 95.38999079009756,
|
| 243 |
+
"items_per_sec_wall": 84.12845368070198,
|
| 244 |
+
"items_per_sec_wall_median": 95.22081950454712,
|
| 245 |
+
"wall": {
|
| 246 |
+
"max_ms": 53.818747168406844,
|
| 247 |
+
"mean_ms": 11.8865848146379,
|
| 248 |
+
"median_ms": 10.501904995180666,
|
| 249 |
+
"min_ms": 10.387809947133064,
|
| 250 |
+
"p95_ms": 10.97815902903676
|
| 251 |
+
}
|
| 252 |
+
},
|
| 253 |
+
"16": {
|
| 254 |
+
"audio_seconds_per_sec_cuda": 1174.1469647216593,
|
| 255 |
+
"audio_seconds_per_sec_cuda_median": 1194.6814835170971,
|
| 256 |
+
"audio_seconds_per_sec_wall": 1173.9429007331944,
|
| 257 |
+
"audio_seconds_per_sec_wall_median": 1194.4542069368163,
|
| 258 |
+
"cuda_events": {
|
| 259 |
+
"max_ms": 172.75180053710938,
|
| 260 |
+
"mean_ms": 136.26914245605468,
|
| 261 |
+
"median_ms": 133.92691040039062,
|
| 262 |
+
"min_ms": 131.3126678466797,
|
| 263 |
+
"p95_ms": 163.50003051757812
|
| 264 |
+
},
|
| 265 |
+
"items_per_sec_cuda": 117.41469647216593,
|
| 266 |
+
"items_per_sec_cuda_median": 119.4681483517097,
|
| 267 |
+
"items_per_sec_wall": 117.39429007331943,
|
| 268 |
+
"items_per_sec_wall_median": 119.44542069368163,
|
| 269 |
+
"wall": {
|
| 270 |
+
"max_ms": 172.77729604393244,
|
| 271 |
+
"mean_ms": 136.29282983019948,
|
| 272 |
+
"median_ms": 133.95239354576916,
|
| 273 |
+
"min_ms": 131.3318640459329,
|
| 274 |
+
"p95_ms": 163.52250101044774
|
| 275 |
+
}
|
| 276 |
+
},
|
| 277 |
+
"8": {
|
| 278 |
+
"audio_seconds_per_sec_cuda": 1305.4346670818657,
|
| 279 |
+
"audio_seconds_per_sec_cuda_median": 1327.4486741223745,
|
| 280 |
+
"audio_seconds_per_sec_wall": 1304.9761567848066,
|
| 281 |
+
"audio_seconds_per_sec_wall_median": 1326.985212046842,
|
| 282 |
+
"cuda_events": {
|
| 283 |
+
"max_ms": 109.47174072265625,
|
| 284 |
+
"mean_ms": 61.28227020263672,
|
| 285 |
+
"median_ms": 60.26598358154297,
|
| 286 |
+
"min_ms": 58.52467346191406,
|
| 287 |
+
"p95_ms": 61.911041259765625
|
| 288 |
+
},
|
| 289 |
+
"items_per_sec_cuda": 130.54346670818657,
|
| 290 |
+
"items_per_sec_cuda_median": 132.74486741223745,
|
| 291 |
+
"items_per_sec_wall": 130.49761567848066,
|
| 292 |
+
"items_per_sec_wall_median": 132.6985212046842,
|
| 293 |
+
"wall": {
|
| 294 |
+
"max_ms": 109.49674621224403,
|
| 295 |
+
"mean_ms": 61.30380205344409,
|
| 296 |
+
"median_ms": 60.28703204356134,
|
| 297 |
+
"min_ms": 58.54333913885057,
|
| 298 |
+
"p95_ms": 61.929865973070264
|
| 299 |
+
}
|
| 300 |
+
}
|
| 301 |
+
}
|
| 302 |
+
},
|
| 303 |
+
"environment": {
|
| 304 |
+
"batch_sizes": [
|
| 305 |
+
1,
|
| 306 |
+
8,
|
| 307 |
+
16
|
| 308 |
+
],
|
| 309 |
+
"cuda_available": true,
|
| 310 |
+
"cuda_device_name": "NVIDIA L4",
|
| 311 |
+
"device": "cuda:0",
|
| 312 |
+
"iters": 50,
|
| 313 |
+
"note": "Synthetic random waveforms; excludes file decode, dataset download, and MTEB result serialization.",
|
| 314 |
+
"platform": "Linux-6.8.12-9-pve-x86_64-with-glibc2.39",
|
| 315 |
+
"python": "3.12.3",
|
| 316 |
+
"seed": 1234,
|
| 317 |
+
"torch": "2.10.0+cu126",
|
| 318 |
+
"warmup": 20
|
| 319 |
+
}
|
| 320 |
+
}
|
manifest.json
CHANGED
|
@@ -4,9 +4,20 @@
|
|
| 4 |
"release_type": "safetensors",
|
| 5 |
"license": "apache-2.0",
|
| 6 |
"parameter_count": 87118774,
|
| 7 |
-
"modalities": [
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
"output_dim": 1280,
|
| 9 |
-
"matryoshka_slices": [
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 10 |
"audio_encoder": {
|
| 11 |
"name": "efficientat_mn20_as_native_merged",
|
| 12 |
"checkpoint": "triembed/checkpoints/mn20_native_merged_aistmix_audioheavy100k175k175k_continue_from_balanced_20260426T143137Z/latest_model.pt",
|
|
@@ -26,7 +37,8 @@
|
|
| 26 |
"768": 0,
|
| 27 |
"512": 0
|
| 28 |
},
|
| 29 |
-
"report": "aist87m_memory_slice_release_report.json"
|
|
|
|
| 30 |
},
|
| 31 |
"files": [
|
| 32 |
{
|
|
@@ -53,6 +65,11 @@
|
|
| 53 |
"path": "aist87m_memory_slice_release_report.json",
|
| 54 |
"size_bytes": 40579,
|
| 55 |
"sha256": "6b5894bd3b8de9e8c99ccd236587fc6f4652399ac95b343ba301478fcdb7c77a"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 56 |
}
|
| 57 |
]
|
| 58 |
}
|
|
|
|
| 4 |
"release_type": "safetensors",
|
| 5 |
"license": "apache-2.0",
|
| 6 |
"parameter_count": 87118774,
|
| 7 |
+
"modalities": [
|
| 8 |
+
"audio",
|
| 9 |
+
"image",
|
| 10 |
+
"speech",
|
| 11 |
+
"text"
|
| 12 |
+
],
|
| 13 |
"output_dim": 1280,
|
| 14 |
+
"matryoshka_slices": [
|
| 15 |
+
1280,
|
| 16 |
+
768,
|
| 17 |
+
512,
|
| 18 |
+
256,
|
| 19 |
+
128
|
| 20 |
+
],
|
| 21 |
"audio_encoder": {
|
| 22 |
"name": "efficientat_mn20_as_native_merged",
|
| 23 |
"checkpoint": "triembed/checkpoints/mn20_native_merged_aistmix_audioheavy100k175k175k_continue_from_balanced_20260426T143137Z/latest_model.pt",
|
|
|
|
| 37 |
"768": 0,
|
| 38 |
"512": 0
|
| 39 |
},
|
| 40 |
+
"report": "aist87m_memory_slice_release_report.json",
|
| 41 |
+
"throughput_report": "aist87m_vs_dual_audio_throughput_l4_20260504.json"
|
| 42 |
},
|
| 43 |
"files": [
|
| 44 |
{
|
|
|
|
| 65 |
"path": "aist87m_memory_slice_release_report.json",
|
| 66 |
"size_bytes": 40579,
|
| 67 |
"sha256": "6b5894bd3b8de9e8c99ccd236587fc6f4652399ac95b343ba301478fcdb7c77a"
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"path": "aist87m_vs_dual_audio_throughput_l4_20260504.json",
|
| 71 |
+
"size_bytes": 13261,
|
| 72 |
+
"sha256": "58bdf1f02d4f5a497f582f4a65978d9a9b6dc37f5ca3905250b0dcdd8ea13817"
|
| 73 |
}
|
| 74 |
]
|
| 75 |
}
|