gemma-e2b-rlcd / reports /cache-bfloat16.json
larkooo's picture
Publish Gemma E2B RLCD with multimodal checkpoint and parallel scoring
53e24ca verified
Raw
History Blame Contribute Delete
7.12 kB
{
"model": "models/gemma-4-e2b-it-4bit",
"status": "cache_numerics_and_timing_probe_not_quality_benchmark",
"results": [
{
"state": "text",
"prefix_tokens": 50,
"suffix_lengths": [
111,
99,
111,
111
],
"cached_serial_vs_uncached": {
"max_logit_delta": 1.15625,
"max_probability_delta": 4.3084425290461535e-05,
"all_winners_match": true
},
"cached_batch_vs_uncached": {
"max_logit_delta": 1.65625,
"max_probability_delta": 5.5105002298472456e-05,
"all_winners_match": true
},
"reversed_batch_vs_original": {
"max_logit_delta": 0.0,
"max_probability_delta": 0.0,
"all_winners_match": true
},
"prefix_unchanged": true,
"branch_batch_sizes": [
4
],
"warm_seconds": {
"uncached": [
0.30837187499855645,
0.29970408300869167,
0.3233241660054773
],
"cached_batched": [
0.38295441700029187,
0.3663168329803739,
0.4002443749923259
]
},
"warm_median_seconds": {
"uncached": 0.30837187499855645,
"cached_batched": 0.38295441700029187
},
"speedup": 0.8052443353808384
},
{
"state": "image",
"prefix_tokens": 295,
"suffix_lengths": [
111,
99,
111,
111
],
"cached_serial_vs_uncached": {
"max_logit_delta": 1.8125,
"max_probability_delta": 0.08184582251343775,
"all_winners_match": true
},
"cached_batch_vs_uncached": {
"max_logit_delta": 1.5,
"max_probability_delta": 0.07954127470411075,
"all_winners_match": true
},
"reversed_batch_vs_original": {
"max_logit_delta": 0.0,
"max_probability_delta": 0.0,
"all_winners_match": true
},
"prefix_unchanged": true,
"branch_batch_sizes": [
4
],
"warm_seconds": {
"uncached": [
2.9452710000041407,
2.088067457982106,
2.104747708013747
],
"cached_batched": [
0.8102280829916708,
0.8053822079964448,
0.8086074579914566
]
},
"warm_median_seconds": {
"uncached": 2.104747708013747,
"cached_batched": 0.8086074579914566
},
"speedup": 2.6029288837402547
},
{
"state": "speech",
"prefix_tokens": 86,
"suffix_lengths": [
111,
99,
111,
111
],
"cached_serial_vs_uncached": {
"max_logit_delta": 0.6875,
"max_probability_delta": 0.1209109506311471,
"all_winners_match": true
},
"cached_batch_vs_uncached": {
"max_logit_delta": 0.6875,
"max_probability_delta": 0.13835689201493084,
"all_winners_match": true
},
"reversed_batch_vs_original": {
"max_logit_delta": 0.0,
"max_probability_delta": 0.0,
"all_winners_match": true
},
"prefix_unchanged": true,
"branch_batch_sizes": [
4
],
"warm_seconds": {
"uncached": [
0.390588584006764,
0.3957693339907564,
0.3943458329886198
],
"cached_batched": [
0.3282808749936521,
0.3258304579940159,
0.3353305420023389
]
},
"warm_median_seconds": {
"uncached": 0.3943458329886198,
"cached_batched": 0.3282808749936521
},
"speedup": 1.2012452233053332
},
{
"state": "video",
"prefix_tokens": 328,
"suffix_lengths": [
111,
99,
111,
111
],
"cached_serial_vs_uncached": {
"max_logit_delta": 0.40625,
"max_probability_delta": 0.016064510780345466,
"all_winners_match": true
},
"cached_batch_vs_uncached": {
"max_logit_delta": 0.5,
"max_probability_delta": 0.022443869382056524,
"all_winners_match": true
},
"reversed_batch_vs_original": {
"max_logit_delta": 0.0,
"max_probability_delta": 0.0,
"all_winners_match": true
},
"prefix_unchanged": true,
"branch_batch_sizes": [
4
],
"warm_seconds": {
"uncached": [
1.817464583000401,
1.9493792499997653,
1.6563971669820603
],
"cached_batched": [
0.7273326250142418,
0.7425714999844786,
0.6501901249866933
]
},
"warm_median_seconds": {
"uncached": 1.817464583000401,
"cached_batched": 0.7273326250142418
},
"speedup": 2.4988080013113856
},
{
"state": "video_speech",
"prefix_tokens": 431,
"suffix_lengths": [
111,
99,
111,
111
],
"cached_serial_vs_uncached": {
"max_logit_delta": 1.1875,
"max_probability_delta": 0.18137001888040638,
"all_winners_match": true
},
"cached_batch_vs_uncached": {
"max_logit_delta": 1.4375,
"max_probability_delta": 0.20775812578824238,
"all_winners_match": true
},
"reversed_batch_vs_original": {
"max_logit_delta": 0.0,
"max_probability_delta": 0.0,
"all_winners_match": true
},
"prefix_unchanged": true,
"branch_batch_sizes": [
4
],
"warm_seconds": {
"uncached": [
1.982235583011061,
2.0264377089915797,
2.0417945420194883
],
"cached_batched": [
0.7379775000154041,
0.7445563329965807,
0.7704525000008289
]
},
"warm_median_seconds": {
"uncached": 2.0264377089915797,
"cached_batched": 0.7445563329965807
},
"speedup": 2.721671442691074
},
{
"state": "long_text",
"prefix_tokens": 1250,
"suffix_lengths": [
111,
99,
111,
111
],
"cached_serial_vs_uncached": {
"max_logit_delta": 1.28125,
"max_probability_delta": 7.581746597207228e-08,
"all_winners_match": true
},
"cached_batch_vs_uncached": {
"max_logit_delta": 1.390625,
"max_probability_delta": 1.2115263459122616e-07,
"all_winners_match": true
},
"reversed_batch_vs_original": {
"max_logit_delta": 0.0,
"max_probability_delta": 0.0,
"all_winners_match": true
},
"prefix_unchanged": true,
"branch_batch_sizes": [
4
],
"warm_seconds": {
"uncached": [
1.2670293750124983,
1.2486699169967324,
1.2255972499842755
],
"cached_batched": [
0.6019650840025861,
0.5762567500059959,
0.592385583993746
]
},
"warm_median_seconds": {
"uncached": 1.2486699169967324,
"cached_batched": 0.592385583993746
},
"speedup": 2.107866819746773
}
]
}