{ "model": "models/gemma-4-e2b-it-4bit", "compute_dtype": "float32", "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": 0.032517433166503906, "max_probability_delta": 4.89707640993764e-07, "all_winners_match": true }, "cached_batch_vs_uncached": { "max_logit_delta": 0.056962013244628906, "max_probability_delta": 5.000832680179812e-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": [ 0.27352558402344584, 0.270969499979401, 0.2627255840052385 ], "cached_batched": [ 0.2238874580070842, 0.22431362498900853, 0.22326229201280512 ] }, "warm_median_seconds": { "uncached": 0.270969499979401, "cached_batched": 0.2238874580070842 }, "speedup": 1.2102933428759866 }, { "state": "image", "prefix_tokens": 295, "suffix_lengths": [ 111, 99, 111, 111 ], "cached_serial_vs_uncached": { "max_logit_delta": 0.04843616485595703, "max_probability_delta": 0.0007440930557797154, "all_winners_match": true }, "cached_batch_vs_uncached": { "max_logit_delta": 0.03708457946777344, "max_probability_delta": 0.001258392859177726, "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": [ 2.799910041008843, 2.575791625014972, 2.678586375026498 ], "cached_batched": [ 0.8003228750021663, 0.774800959014101, 0.7754046249901876 ] }, "warm_median_seconds": { "uncached": 2.678586375026498, "cached_batched": 0.7754046249901876 }, "speedup": 3.4544369335692244 }, { "state": "speech", "prefix_tokens": 86, "suffix_lengths": [ 111, 99, 111, 111 ], "cached_serial_vs_uncached": { "max_logit_delta": 0.08770561218261719, "max_probability_delta": 0.013604034682185018, "all_winners_match": true }, "cached_batch_vs_uncached": { "max_logit_delta": 0.03138256072998047, "max_probability_delta": 0.002520692809385161, "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": [ 0.29824562498833984, 0.3047535420046188, 0.304851624998264 ], "cached_batched": [ 0.2058587910141796, 0.1957708330010064, 0.20528816597652622 ] }, "warm_median_seconds": { "uncached": 0.3047535420046188, "cached_batched": 0.20528816597652622 }, "speedup": 1.4845158782287822 }, { "state": "video", "prefix_tokens": 328, "suffix_lengths": [ 111, 99, 111, 111 ], "cached_serial_vs_uncached": { "max_logit_delta": 0.007994651794433594, "max_probability_delta": 0.00047290616548957554, "all_winners_match": true }, "cached_batch_vs_uncached": { "max_logit_delta": 0.044986724853515625, "max_probability_delta": 0.003810715236252593, "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": [ 3.470609209005488, 4.113289792003343, 3.7244844590022694 ], "cached_batched": [ 1.141944707982475, 1.2892172499850858, 1.1141800410114229 ] }, "warm_median_seconds": { "uncached": 3.7244844590022694, "cached_batched": 1.141944707982475 }, "speedup": 3.2615278419061844 }, { "state": "video_speech", "prefix_tokens": 431, "suffix_lengths": [ 111, 99, 111, 111 ], "cached_serial_vs_uncached": { "max_logit_delta": 0.05509662628173828, "max_probability_delta": 0.005538062741424418, "all_winners_match": true }, "cached_batch_vs_uncached": { "max_logit_delta": 0.03436756134033203, "max_probability_delta": 0.0013212921945262024, "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": [ 4.29815650000819, 3.8582746669999324, 3.3546597079839557 ], "cached_batched": [ 1.3747120829939377, 1.1818989999883343, 1.4414640410104766 ] }, "warm_median_seconds": { "uncached": 3.8582746669999324, "cached_batched": 1.3747120829939377 }, "speedup": 2.806605626537544 }, { "state": "long_text", "prefix_tokens": 1250, "suffix_lengths": [ 111, 99, 111, 111 ], "cached_serial_vs_uncached": { "max_logit_delta": 0.028132915496826172, "max_probability_delta": 1.1824594636777874e-09, "all_winners_match": true }, "cached_batch_vs_uncached": { "max_logit_delta": 0.0856776237487793, "max_probability_delta": 3.4840531570878852e-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": [ 3.0424842079810333, 2.5938880419998895, 2.5387540830124635 ], "cached_batched": [ 1.0249322910094634, 0.8549119170056656, 0.8327442500158213 ] }, "warm_median_seconds": { "uncached": 2.5938880419998895, "cached_batched": 0.8549119170056656 }, "speedup": 3.034099759756536 } ] }