{ "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": 0.05619525909423828, "max_probability_delta": 1.9645769933350452e-07, "all_winners_match": true }, "cached_batch_vs_uncached": { "max_logit_delta": 0.06667137145996094, "max_probability_delta": 2.2873066432471377e-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": [ 0.2668775000202004, 0.26092958298977464, 0.2707252079853788 ], "cached_batched": [ 0.40814466698793694, 0.389646374998847, 0.4031292090076022 ] }, "warm_median_seconds": { "uncached": 0.2668775000202004, "cached_batched": 0.4031292090076022 }, "speedup": 0.6620147934137107 }, { "state": "image", "prefix_tokens": 295, "suffix_lengths": [ 111, 99, 111, 111 ], "cached_serial_vs_uncached": { "max_logit_delta": 0.029593229293823242, "max_probability_delta": 0.00047943623264679047, "all_winners_match": true }, "cached_batch_vs_uncached": { "max_logit_delta": 0.039635658264160156, "max_probability_delta": 0.0025423919720055643, "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": [ 2.447480083006667, 2.5804249589855317, 2.4646622499858495 ], "cached_batched": [ 0.9259148749988526, 0.9439381250122096, 0.9788283329980914 ] }, "warm_median_seconds": { "uncached": 2.4646622499858495, "cached_batched": 0.9439381250122096 }, "speedup": 2.61104216969091 }, { "state": "speech", "prefix_tokens": 86, "suffix_lengths": [ 111, 99, 111, 111 ], "cached_serial_vs_uncached": { "max_logit_delta": 0.07565116882324219, "max_probability_delta": 0.010469679977858998, "all_winners_match": true }, "cached_batch_vs_uncached": { "max_logit_delta": 0.05370330810546875, "max_probability_delta": 0.0077296466941121356, "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": [ 0.3987604579888284, 0.4338792499911506, 0.4244323339953553 ], "cached_batched": [ 0.45298162501421757, 0.45572150000953116, 0.45018458299455233 ] }, "warm_median_seconds": { "uncached": 0.4244323339953553, "cached_batched": 0.45298162501421757 }, "speedup": 0.9369747260322839 }, { "state": "video", "prefix_tokens": 328, "suffix_lengths": [ 111, 99, 111, 111 ], "cached_serial_vs_uncached": { "max_logit_delta": 0.020721912384033203, "max_probability_delta": 0.0023643622739676073, "all_winners_match": true }, "cached_batch_vs_uncached": { "max_logit_delta": 0.04329872131347656, "max_probability_delta": 0.003217125559550421, "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": [ 2.826576583989663, 2.8505677079956513, 2.8707061669847462 ], "cached_batched": [ 1.1707769579952583, 1.1337155420042109, 1.001873125002021 ] }, "warm_median_seconds": { "uncached": 2.8505677079956513, "cached_batched": 1.1337155420042109 }, "speedup": 2.5143588513890758 }, { "state": "video_speech", "prefix_tokens": 431, "suffix_lengths": [ 111, 99, 111, 111 ], "cached_serial_vs_uncached": { "max_logit_delta": 0.05681800842285156, "max_probability_delta": 0.003680399076730595, "all_winners_match": true }, "cached_batch_vs_uncached": { "max_logit_delta": 0.04210376739501953, "max_probability_delta": 0.0027342631865970995, "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": [ 2.9628333750006277, 3.1171925840026233, 2.7277989579888526 ], "cached_batched": [ 1.1268984170164913, 1.1107918329944368, 1.1665158329997212 ] }, "warm_median_seconds": { "uncached": 2.9628333750006277, "cached_batched": 1.1268984170164913 }, "speedup": 2.629192951432879 }, { "state": "long_text", "prefix_tokens": 1250, "suffix_lengths": [ 111, 99, 111, 111 ], "cached_serial_vs_uncached": { "max_logit_delta": 0.04975175857543945, "max_probability_delta": 1.6751281473048607e-09, "all_winners_match": true }, "cached_batch_vs_uncached": { "max_logit_delta": 0.0802912712097168, "max_probability_delta": 3.9548890828911e-09, "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": [ 2.120250792009756, 2.2143384579976555, 2.0523968339839485 ], "cached_batched": [ 1.1375709579733666, 0.9040895829966757, 1.0827704170078505 ] }, "warm_median_seconds": { "uncached": 2.120250792009756, "cached_batched": 1.0827704170078505 }, "speedup": 1.9581720729579033 } ], "compute_dtype": "float32" }