{ "utc": "2026-10-30T22:10:40.353534+00:00", "status": "ALL12_H100_EXLLAMA_TRIALS_RECOVERED_AUDITED_OUTPUTS_REVIEWED", "model": "GLM-5.3-Flash EXL3 sc_3.10bpw", "model_revision": "b570edf655a2bb00beca23e18d3b98724bfc2db5", "engine": "ExLlamaV3-1.6.5-gpuembed1", "runtime_container_image": "sha256:f10e766256cbd73a543de2ed90948d165935f30bbb2edab526095063f8a47e07", "compiler": "GCC12.2 inside new derived container only", "python": "3.12.15", "hardware": "H100 PCIe 80GB x2, PHB interconnect, existing 350W power limit per GPU", "communication_environment": { "NCCL_P2P_DISABLE": "1" }, "observed_NCCL_transport": "SHM/direct/direct", "initial_failures_preserved": [ "capacity-preload-stall", "default-NCCL-tiny-allreduce-timeout", "cumem-disabled-tiny-allreduce-timeout", "missing-C-compiler-at-warmup" ], "performance_integrity_pass": true, "answer_quality_acceptance_pass": false, "main_trials": 12, "groups": [ { "actual_input_tokens": 8193, "n_each": 3, "H100_ExLlama": { "off": { "actual_input_tokens": 8193, "n": 3, "all_conservative_intertoken_tps_above50": true, "generation_tps_native": { "values": [ 52.3294371283103, 52.49864342744915, 52.56531797768627 ], "median": 52.49864342744915, "min": 52.3294371283103, "max": 52.56531797768627 }, "generation_tps_intertoken": { "values": [ 52.25414297416885, 52.44676729758014, 52.50855197555054 ], "median": 52.44676729758014, "min": 52.25414297416885, "max": 52.50855197555054 }, "input_seconds_native": { "values": [ 3.7245123386383057, 3.0325634479522705, 3.031067132949829 ], "median": 3.0325634479522705, "min": 3.031067132949829, "max": 3.7245123386383057 }, "input_tps_native": { "values": [ 2199.7510694233297, 2701.6747186385496, 2703.0084259554446 ], "median": 2701.6747186385496, "min": 2199.7510694233297, "max": 2703.0084259554446 }, "first_generated_seconds": { "values": [ 3.7546712030889466, 3.0554310180014, 3.0524383690208197 ], "median": 3.0554310180014, "min": 3.0524383690208197, "max": 3.7546712030889466 }, "first_visible_seconds": { "values": [ 3.7546712030889466, 3.0741681560175493, 3.0712107800645754 ], "median": 3.0741681560175493, "min": 3.0712107800645754, "max": 3.7546712030889466 }, "first_answer_visible_seconds": { "values": [ 3.946284916019067, 3.0741681560175493, 3.0712107800645754 ], "median": 3.0741681560175493, "min": 3.0712107800645754, "max": 3.946284916019067 }, "end_to_end_seconds": { "values": [ 17.147184645058587, 22.31508721492719, 20.653036167030223 ], "median": 20.653036167030223, "min": 17.147184645058587, "max": 22.31508721492719 } }, "on": { "actual_input_tokens": 8193, "n": 3, "all_conservative_intertoken_tps_above50": true, "generation_tps_native": { "values": [ 62.69125770784577, 71.25922708831844, 75.4603340085201 ], "median": 71.25922708831844, "min": 62.69125770784577, "max": 75.4603340085201 }, "generation_tps_intertoken": { "values": [ 62.61190168543078, 71.18651359128955, 75.38657024116084 ], "median": 71.18651359128955, "min": 62.61190168543078, "max": 75.38657024116084 }, "input_seconds_native": { "values": [ 3.824756383895874, 3.116899251937866, 3.119197368621826 ], "median": 3.119197368621826, "min": 3.116899251937866, "max": 3.824756383895874 }, "input_tps_native": { "values": [ 2142.09721552374, 2628.5738927577386, 2626.637250473176 ], "median": 2626.637250473176, "min": 2142.09721552374, "max": 2628.5738927577386 }, "first_generated_seconds": { "values": [ 4.515950244036503, 3.15280508401338, 3.1550411159405485 ], "median": 3.1550411159405485, "min": 3.15280508401338, "max": 4.515950244036503 }, "first_visible_seconds": { "values": [ 4.515950244036503, 3.1528311029542238, 3.1550662230001763 ], "median": 3.1550662230001763, "min": 3.1528311029542238, "max": 4.515950244036503 }, "first_answer_visible_seconds": { "values": [ 5.725612033973448, 3.1528311029542238, 3.1550662230001763 ], "median": 3.1550662230001763, "min": 3.1528311029542238, "max": 5.725612033973448 }, "end_to_end_seconds": { "values": [ 16.570731892017648, 16.875756316934712, 16.68231029296294 ], "median": 16.68231029296294, "min": 16.570731892017648, "max": 16.875756316934712 }, "draft_acceptance_ratio": { "values": [ 0.47692307692307695, 0.4634146341463415, 0.5045045045045045 ], "median": 0.47692307692307695, "min": 0.4634146341463415, "max": 0.5045045045045045 } } }, "B200_ExLlama": { "off": { "actual_input_tokens": 8193, "n": 3, "all_conservative_intertoken_tps_above50": true, "generation_tps_native": { "values": [ 60.09055744850449, 59.4433888937625, 59.524057443092495 ], "median": 59.524057443092495, "min": 59.4433888937625, "max": 60.09055744850449 }, "generation_tps_intertoken": { "values": [ 60.015066798443556, 59.377778089243996, 59.456029948871816 ], "median": 59.456029948871816, "min": 59.377778089243996, "max": 60.015066798443556 }, "input_seconds_native": { "values": [ 2.478013515472412, 2.092078685760498, 2.0884130001068115 ], "median": 2.092078685760498, "min": 2.0884130001068115, "max": 2.478013515472412 }, "input_tps_native": { "values": [ 3306.2773664646757, 3916.200693484785, 3923.074602380358 ], "median": 3916.200693484785, "min": 3306.2773664646757, "max": 3923.074602380358 }, "first_generated_seconds": { "values": [ 2.511534914956428, 2.1108196099521592, 2.1073353299871087 ], "median": 2.1108196099521592, "min": 2.1073353299871087, "max": 2.511534914956428 }, "first_visible_seconds": { "values": [ 2.511534914956428, 2.1271576220169663, 2.12368782190606 ], "median": 2.1271576220169663, "min": 2.12368782190606, "max": 2.511534914956428 }, "first_answer_visible_seconds": { "values": [ 2.6757730098906904, 2.1271576220169663, 2.12368782190606 ], "median": 2.1271576220169663, "min": 2.12368782190606, "max": 2.6757730098906904 }, "end_to_end_seconds": { "values": [ 15.784663100959733, 17.33802134799771, 16.793299278942868 ], "median": 16.793299278942868, "min": 15.784663100959733, "max": 17.33802134799771 } }, "on": { "actual_input_tokens": 8193, "n": 3, "all_conservative_intertoken_tps_above50": true, "generation_tps_native": { "values": [ 82.89585427095253, 95.55501805046923, 100.92978586160334 ], "median": 95.55501805046923, "min": 82.89585427095253, "max": 100.92978586160334 }, "generation_tps_intertoken": { "values": [ 82.79513026819316, 95.46170260315432, 100.81390780665316 ], "median": 95.46170260315432, "min": 82.79513026819316, "max": 100.81390780665316 }, "input_seconds_native": { "values": [ 2.5195040702819824, 2.134533405303955, 2.135387897491455 ], "median": 2.135387897491455, "min": 2.134533405303955, "max": 2.5195040702819824 }, "input_tps_native": { "values": [ 3251.830428312442, 3838.3095713760104, 3836.7736417466444 ], "median": 3836.7736417466444, "min": 3251.830428312442, "max": 3838.3095713760104 }, "first_generated_seconds": { "values": [ 3.1026840769918635, 2.1630470579257235, 2.1635261320043355 ], "median": 2.1635261320043355, "min": 2.1630470579257235, "max": 3.1026840769918635 }, "first_visible_seconds": { "values": [ 3.1026840769918635, 2.1630665080156177, 2.16354475193657 ], "median": 2.16354475193657, "min": 2.1630665080156177, "max": 3.1026840769918635 }, "first_answer_visible_seconds": { "values": [ 3.919887642026879, 2.1630665080156177, 2.16354475193657 ], "median": 2.16354475193657, "min": 2.1630665080156177, "max": 3.919887642026879 }, "end_to_end_seconds": { "values": [ 12.449465634999797, 12.85473516897764, 10.768680646899156 ], "median": 12.449465634999797, "min": 10.768680646899156, "max": 12.85473516897764 }, "draft_acceptance_ratio": { "values": [ 0.4929718875502008, 0.49514563106796117, 0.5411646586345381 ], "median": 0.49514563106796117, "min": 0.4929718875502008, "max": 0.5411646586345381 } } }, "H100_llamacpp": { "input_tokens": 8193, "repetitions_each": 3, "off": { "actual_input_tokens": 8193, "repetitions": 3, "generation_tokens_per_second": { "all": [ 41.28226747667661, 51.22410812255903, 51.153848406658035 ], "median": 51.153848406658035, "min": 41.28226747667661, "max": 51.22410812255903 }, "first_streamed_generated_text_seconds": { "all": [ 22.92088064807467, 22.474210544954985, 22.475983759970404 ], "median": 22.475983759970404, "min": 22.474210544954985, "max": 22.92088064807467 }, "first_answer_text_seconds": { "all": [ 27.273759820032865, 22.474210544954985, 22.475983759970404 ], "median": 22.475983759970404, "min": 22.474210544954985, "max": 27.273759820032865 }, "input_processing_seconds": { "all": [ 22.68219, 22.430748, 22.433433 ], "median": 22.433433, "min": 22.430748, "max": 22.68219 }, "end_to_end_seconds": { "all": [ 41.30697966204025, 41.69950583600439, 40.37881925900001 ], "median": 41.30697966204025, "min": 40.37881925900001, "max": 41.69950583600439 }, "capped_trials": [], "draft_acceptance_fraction": null }, "on": { "actual_input_tokens": 8193, "repetitions": 3, "generation_tokens_per_second": { "all": [ 61.51962611014726, 62.27434301329048, 62.77850289778328 ], "median": 62.27434301329048, "min": 61.51962611014726, "max": 62.77850289778328 }, "first_streamed_generated_text_seconds": { "all": [ 24.887945767026395, 24.628569833934307, 24.631712430040352 ], "median": 24.631712430040352, "min": 24.628569833934307, "max": 24.887945767026395 }, "first_answer_text_seconds": { "all": [ 26.091206192970276, 24.628569833934307, 24.631712430040352 ], "median": 24.631712430040352, "min": 24.628569833934307, "max": 26.091206192970276 }, "input_processing_seconds": { "all": [ 24.648375, 24.562103999999998, 24.565319 ], "median": 24.565319, "min": 24.562103999999998, "max": 24.648375 }, "end_to_end_seconds": { "all": [ 39.339272454963066, 41.010106497909874, 40.88068579498213 ], "median": 40.88068579498213, "min": 39.339272454963066, "max": 41.010106497909874 }, "capped_trials": [ "ctx8192-r2", "ctx8192-r3" ], "draft_acceptance_fraction": { "all": [ 0.5265700483091788, 0.5370212765957447, 0.5474137931034483 ], "median": 0.5370212765957447, "min": 0.5265700483091788, "max": 0.5474137931034483 } }, "generation_median_ratio": 1.217393118074476, "generation_median_change_percent": 21.739311807447592, "first_streamed_text_median_change_seconds": 2.155728670069948, "on_all_trial_means_above_50": true }, "H100_ExLlama_MTP_speed_ratio": 1.3573098449973304, "H100_to_B200_ExLlama_generation_ratio": { "off": 0.8821101466525906, "on": 0.7457075628246479 } }, { "actual_input_tokens": 32768, "n_each": 3, "H100_ExLlama": { "off": { "actual_input_tokens": 32768, "n": 3, "all_conservative_intertoken_tps_above50": true, "generation_tps_native": { "values": [ 52.21566454622748, 52.186378904911926, 52.27313594414483 ], "median": 52.21566454622748, "min": 52.186378904911926, "max": 52.27313594414483 }, "generation_tps_intertoken": { "values": [ 52.162437569116136, 52.13150363792464, 52.21805360911412 ], "median": 52.162437569116136, "min": 52.13150363792464, "max": 52.21805360911412 }, "input_seconds_native": { "values": [ 12.836126565933228, 12.055091142654419, 12.045980453491211 ], "median": 12.055091142654419, "min": 12.045980453491211, "max": 12.836126565933228 }, "input_tps_native": { "values": [ 2552.7950220563803, 2718.187661315748, 2720.243497531416 ], "median": 2718.187661315748, "min": 2552.7950220563803, "max": 2720.243497531416 }, "first_generated_seconds": { "values": [ 12.858338313060813, 12.077334010973573, 12.068147322977893 ], "median": 12.077334010973573, "min": 12.068147322977893, "max": 12.858338313060813 }, "first_visible_seconds": { "values": [ 12.858338313060813, 12.096278497949243, 12.068147322977893 ], "median": 12.096278497949243, "min": 12.068147322977893, "max": 12.858338313060813 }, "first_answer_visible_seconds": { "values": [ 13.985714378068224, 12.096278497949243, 13.875998331932351 ], "median": 13.875998331932351, "min": 12.096278497949243, "max": 13.985714378068224 }, "end_to_end_seconds": { "values": [ 31.629685522057116, 30.286028780974448, 30.207308548036963 ], "median": 30.286028780974448, "min": 30.207308548036963, "max": 31.629685522057116 } }, "on": { "actual_input_tokens": 32768, "n": 3, "all_conservative_intertoken_tps_above50": true, "generation_tps_native": { "values": [ 75.20312987333955, 74.32928037543233, 75.94182420650134 ], "median": 75.20312987333955, "min": 74.32928037543233, "max": 75.94182420650134 }, "generation_tps_intertoken": { "values": [ 75.121916990107, 74.25472743623932, 75.85403019007764 ], "median": 75.121916990107, "min": 74.25472743623932, "max": 75.85403019007764 }, "input_seconds_native": { "values": [ 13.206957578659058, 12.403007745742798, 12.410313367843628 ], "median": 12.410313367843628, "min": 12.403007745742798, "max": 13.206957578659058 }, "input_tps_native": { "values": [ 2481.1164724985083, 2641.9398158682334, 2640.3845760176523 ], "median": 2640.3845760176523, "min": 2481.1164724985083, "max": 2641.9398158682334 }, "first_generated_seconds": { "values": [ 13.243030342971906, 12.439681946998462, 12.446683499962091 ], "median": 12.446683499962091, "min": 12.439681946998462, "max": 13.243030342971906 }, "first_visible_seconds": { "values": [ 13.243030342971906, 12.439706002944149, 12.446683499962091 ], "median": 12.446683499962091, "min": 12.439706002944149, "max": 13.243030342971906 }, "first_answer_visible_seconds": { "values": [ 13.888618264929391, 12.439706002944149, 13.65313025994692 ], "median": 13.65313025994692, "min": 12.439706002944149, "max": 13.888618264929391 }, "end_to_end_seconds": { "values": [ 25.52739283500705, 25.82369013596326, 23.808734862948768 ], "median": 25.52739283500705, "min": 23.808734862948768, "max": 25.82369013596326 }, "draft_acceptance_ratio": { "values": [ 0.5100182149362478, 0.5058922558922558, 0.5172074729596854 ], "median": 0.5100182149362478, "min": 0.5058922558922558, "max": 0.5172074729596854 } } }, "B200_ExLlama": { "off": { "actual_input_tokens": 32768, "n": 3, "all_conservative_intertoken_tps_above50": true, "generation_tps_native": { "values": [ 59.67992378462389, 60.15170206566067, 58.888886431651 ], "median": 59.67992378462389, "min": 58.888886431651, "max": 60.15170206566067 }, "generation_tps_intertoken": { "values": [ 59.614699824203534, 60.08988120022834, 58.82786167887209 ], "median": 59.614699824203534, "min": 58.82786167887209, "max": 60.08988120022834 }, "input_seconds_native": { "values": [ 8.95195460319519, 8.263920545578003, 8.264732360839844 ], "median": 8.264732360839844, "min": 8.263920545578003, "max": 8.95195460319519 }, "input_tps_native": { "values": [ 3660.4296438572455, 3965.188171797471, 3964.798685467679 ], "median": 3964.798685467679, "min": 3660.4296438572455, "max": 3965.188171797471 }, "first_generated_seconds": { "values": [ 8.971125897020102, 8.28313649003394, 8.28419387107715 ], "median": 8.28419387107715, "min": 8.28313649003394, "max": 8.971125897020102 }, "first_visible_seconds": { "values": [ 8.971125897020102, 8.299386703991331, 8.30083858803846 ], "median": 8.30083858803846, "min": 8.299386703991331, "max": 8.971125897020102 }, "first_answer_visible_seconds": { "values": [ 9.810961136012338, 8.299386703991331, 8.30083858803846 ], "median": 8.30083858803846, "min": 8.299386703991331, "max": 9.810961136012338 }, "end_to_end_seconds": { "values": [ 24.287577643990517, 24.445283739012666, 24.656483210041188 ], "median": 24.445283739012666, "min": 24.287577643990517, "max": 24.656483210041188 } }, "on": { "actual_input_tokens": 32768, "n": 3, "all_conservative_intertoken_tps_above50": true, "generation_tps_native": { "values": [ 93.06145845088606, 93.02216469748399, 92.72427426220128 ], "median": 93.02216469748399, "min": 92.72427426220128, "max": 93.06145845088606 }, "generation_tps_intertoken": { "values": [ 92.95449125726437, 92.9313227397716, 92.62975410495133 ], "median": 92.9313227397716, "min": 92.62975410495133, "max": 92.95449125726437 }, "input_seconds_native": { "values": [ 9.163030624389648, 8.46372389793396, 8.459462881088257 ], "median": 8.46372389793396, "min": 8.459462881088257, "max": 9.163030624389648 }, "input_tps_native": { "values": [ 3576.1094056348506, 3871.581870481248, 3873.5319795840987 ], "median": 3871.581870481248, "min": 3576.1094056348506, "max": 3873.5319795840987 }, "first_generated_seconds": { "values": [ 9.192164542968385, 8.492621793993749, 8.48864426300861 ], "median": 8.492621793993749, "min": 8.48864426300861, "max": 9.192164542968385 }, "first_visible_seconds": { "values": [ 9.192164542968385, 8.492649144027382, 8.488666923018172 ], "median": 8.492649144027382, "min": 8.488666923018172, "max": 9.192164542968385 }, "first_answer_visible_seconds": { "values": [ 9.907278175000101, 8.492649144027382, 8.488666923018172 ], "median": 8.492649144027382, "min": 8.488666923018172, "max": 9.907278175000101 }, "end_to_end_seconds": { "values": [ 18.51600027899258, 19.47699537104927, 19.044399930979125 ], "median": 19.044399930979125, "min": 18.51600027899258, "max": 19.47699537104927 }, "draft_acceptance_ratio": { "values": [ 0.4767225325884544, 0.4755134281200632, 0.4720853858784893 ], "median": 0.4755134281200632, "min": 0.4720853858784893, "max": 0.4767225325884544 } } }, "H100_llamacpp": { "input_tokens": 32768, "repetitions_each": 3, "off": { "actual_input_tokens": 32768, "repetitions": 3, "generation_tokens_per_second": { "all": [ 49.71899775054341, 49.72888671075154, 49.767699383269026 ], "median": 49.72888671075154, "min": 49.71899775054341, "max": 49.767699383269026 }, "first_streamed_generated_text_seconds": { "all": [ 86.56187437300105, 86.54652248800267, 86.54690774192568 ], "median": 86.54690774192568, "min": 86.54652248800267, "max": 86.56187437300105 }, "first_answer_text_seconds": { "all": [ 88.69535350007936, 86.75351399497595, 87.07444424391724 ], "median": 87.07444424391724, "min": 86.75351399497595, "max": 88.69535350007936 }, "input_processing_seconds": { "all": [ 86.504044, 86.488305, 86.48918300000001 ], "median": 86.48918300000001, "min": 86.488305, "max": 86.504044 }, "end_to_end_seconds": { "all": [ 107.13822314806748, 103.05686300690286, 101.47718459099997 ], "median": 103.05686300690286, "min": 101.47718459099997, "max": 107.13822314806748 }, "capped_trials": [ "ctx32768-r1" ], "draft_acceptance_fraction": null }, "on": { "actual_input_tokens": 32768, "repetitions": 3, "generation_tokens_per_second": { "all": [ 55.7749175125499, 53.98533985476238, 57.01239519928313 ], "median": 55.7749175125499, "min": 53.98533985476238, "max": 57.01239519928313 }, "first_streamed_generated_text_seconds": { "all": [ 94.31905742699746, 94.38870209804736, 94.3602513789665 ], "median": 94.3602513789665, "min": 94.31905742699746, "max": 94.38870209804736 }, "first_answer_text_seconds": { "all": [ 96.27599957806524, 95.75533016200643, 94.7251296370523 ], "median": 95.75533016200643, "min": 94.7251296370523, "max": 96.27599957806524 }, "input_processing_seconds": { "all": [ 94.26143300000001, 94.330284, 94.30095299999999 ], "median": 94.30095299999999, "min": 94.26143300000001, "max": 94.330284 }, "end_to_end_seconds": { "all": [ 112.66147581802215, 110.22721290506888, 108.27052422100678 ], "median": 110.22721290506888, "min": 108.27052422100678, "max": 112.66147581802215 }, "capped_trials": [ "ctx32768-r1" ], "draft_acceptance_fraction": { "all": [ 0.5132450331125827, 0.48799231508165225, 0.5401540154015402 ], "median": 0.5132450331125827, "min": 0.48799231508165225, "max": 0.5401540154015402 } }, "generation_median_ratio": 1.1215798543201891, "generation_median_change_percent": 12.157985432018915, "first_streamed_text_median_change_seconds": 7.813343637040816, "on_all_trial_means_above_50": true }, "H100_ExLlama_MTP_speed_ratio": 1.4401534991643967, "H100_to_B200_ExLlama_generation_ratio": { "off": 0.8749928746254999, "on": 0.8083594936065324 } } ], "records": { "off": [ { "case": "ctx8192-r1", "actual_input_tokens": 8193, "reference_input_tokens": 8192, "input_content_sha256": "fe7772257f2e32cecb8ed3052c1bc076e77e0c0b68e97fcb11cffb2f0035660c", "render_sha256": "2785e85f267cc7917263671a8b368ee917ff67ad3bc7e1a050abcfb98c4ab4dc", "generated_tokens": 695, "generation_seconds_native": 13.281243562698364, "generation_tps_native": 52.3294371283103, "generation_tps_intertoken": 52.25414297416885, "input_seconds_native": 3.7245123386383057, "input_tps_native": 2199.7510694233297, "first_generated_seconds": 3.7546712030889466, "first_visible_seconds": 3.7546712030889466, "first_answer_visible_seconds": 3.946284916019067, "end_to_end_seconds": 17.147184645058587, "finish_reason": "stop_token", "cached_tokens": 0, "context_shift_or_requeue": false, "performance_integrity_checks_pass": true, "answer_quality_pass": false, "answer_quality_review": "FULL_OUTPUT_REVIEWED_BY_CODEX_SEPARATE_FROM_SPEED", "output_sha256": "f43dbdbe4bab218fc16ed0ee8117ef797c34430528484559b5c980b320fa17b9", "quality_findings": [ "開始待ちを「最大80秒超」と記述したが、入力の最大値は80秒。", "入力処理時間を最初のトークン出力までと定義し、TTFTとprefillを区別していない。" ] }, { "case": "ctx32768-r1", "actual_input_tokens": 32768, "reference_input_tokens": 32768, "input_content_sha256": "5861ad9ee43affadb6a889494a29ef9097c485e5671850050a8e9f24c5214014", "render_sha256": "8741d3b84a76b641a93767e66d9a578272d32b01885611d215ce8cfcac71a546", "generated_tokens": 981, "generation_seconds_native": 18.787465572357178, "generation_tps_native": 52.21566454622748, "generation_tps_intertoken": 52.162437569116136, "input_seconds_native": 12.836126565933228, "input_tps_native": 2552.7950220563803, "first_generated_seconds": 12.858338313060813, "first_visible_seconds": 12.858338313060813, "first_answer_visible_seconds": 13.985714378068224, "end_to_end_seconds": 31.629685522057116, "finish_reason": "stop_token", "cached_tokens": 0, "context_shift_or_requeue": false, "performance_integrity_checks_pass": true, "answer_quality_pass": false, "answer_quality_review": "FULL_OUTPUT_REVIEWED_BY_CODEX_SEPARATE_FROM_SPEED", "output_sha256": "b1fe8f6e1d76bcb6a7293c6967d779f88e519b54a49b1a19a9c20362afb32f9b", "quality_findings": [ "入力処理の評価を最初のトークンまでの時間から待ちを除く手順で説明し、純粋なprefillと初回生成を分離していない。", "「変動が2倍」などの判定基準は比較対象・根拠が十分に定義されていない。" ] }, { "case": "ctx32768-r2", "actual_input_tokens": 32768, "reference_input_tokens": 32768, "input_content_sha256": "df87c9d1b8895eaf28156c4236df58438ee505413bfa71c5071520e755fab152", "render_sha256": "8741d3b84a76b641a93767e66d9a578272d32b01885611d215ce8cfcac71a546", "generated_tokens": 951, "generation_seconds_native": 18.223145961761475, "generation_tps_native": 52.186378904911926, "generation_tps_intertoken": 52.13150363792464, "input_seconds_native": 12.055091142654419, "input_tps_native": 2718.187661315748, "first_generated_seconds": 12.077334010973573, "first_visible_seconds": 12.096278497949243, "first_answer_visible_seconds": 12.096278497949243, "end_to_end_seconds": 30.286028780974448, "finish_reason": "stop_token", "cached_tokens": 0, "context_shift_or_requeue": false, "performance_integrity_checks_pass": true, "answer_quality_pass": false, "answer_quality_review": "FULL_OUTPUT_REVIEWED_BY_CODEX_SEPARATE_FROM_SPEED", "output_sha256": "2d67ff97071d05b4576eb2cb067fc142ad8e2dc12d331ad8efe94007490c7c98", "quality_findings": [ "入力長が10周期（730トークン）で周回するとしたが、実際は記録00411でリセットする。", "入力長・同時利用数と待ち時間の因果的な傾向を架空の混在記録から断定している。", "VRAM80%・待ち10秒などの基準はこの記録から検証されていない。" ] }, { "case": "ctx8192-r2", "actual_input_tokens": 8193, "reference_input_tokens": 8192, "input_content_sha256": "227fa296fc7b644c484fc40877bdb35e587d70539a7bd314d37a43409cdf2dbb", "render_sha256": "2785e85f267cc7917263671a8b368ee917ff67ad3bc7e1a050abcfb98c4ab4dc", "generated_tokens": 1012, "generation_seconds_native": 19.27668857574463, "generation_tps_native": 52.49864342744915, "generation_tps_intertoken": 52.44676729758014, "input_seconds_native": 3.0325634479522705, "input_tps_native": 2701.6747186385496, "first_generated_seconds": 3.0554310180014, "first_visible_seconds": 3.0741681560175493, "first_answer_visible_seconds": 3.0741681560175493, "end_to_end_seconds": 22.31508721492719, "finish_reason": "stop_token", "cached_tokens": 0, "context_shift_or_requeue": false, "performance_integrity_checks_pass": true, "answer_quality_pass": false, "answer_quality_review": "FULL_OUTPUT_REVIEWED_BY_CODEX_SEPARATE_FROM_SPEED", "output_sha256": "b7eea925dfd4fd39b887dfb3c51e7cb57e97fb6db47761e10d014a98d23d2c90", "quality_findings": [ "開始待ち0〜80秒は誤りで、最小値は1秒。4件同時の最大74秒も誤りで、実際は78秒。", "TTFTがprefill時間に相当するとしており、キューや初回生成を区別していない。" ] }, { "case": "ctx8192-r3", "actual_input_tokens": 8193, "reference_input_tokens": 8192, "input_content_sha256": "e50d4d80b003558993083950a30517f1b8aa99cb95c542b50b33782237d011f6", "render_sha256": "2785e85f267cc7917263671a8b368ee917ff67ad3bc7e1a050abcfb98c4ab4dc", "generated_tokens": 926, "generation_seconds_native": 17.616178035736084, "generation_tps_native": 52.56531797768627, "generation_tps_intertoken": 52.50855197555054, "input_seconds_native": 3.031067132949829, "input_tps_native": 2703.0084259554446, "first_generated_seconds": 3.0524383690208197, "first_visible_seconds": 3.0712107800645754, "first_answer_visible_seconds": 3.0712107800645754, "end_to_end_seconds": 20.653036167030223, "finish_reason": "stop_token", "cached_tokens": 0, "context_shift_or_requeue": false, "performance_integrity_checks_pass": true, "answer_quality_pass": false, "answer_quality_review": "FULL_OUTPUT_REVIEWED_BY_CODEX_SEPARATE_FROM_SPEED", "output_sha256": "33b2549cf97467d41cead73011bdc2dee7c7881b0fbd9d2f6732964a2ede2a43", "quality_findings": [ "入力長と開始待ちが並行して増えるとしたが、待ち時間は循環しており、因果的な単調増加は示されない。", "入力処理測定でTTFTと記録の開始待ちを対応させ、prefillとの区別が不十分。", "VRAM90%や変動係数0.3の判定基準は本記録から検証されていない。" ] }, { "case": "ctx32768-r3", "actual_input_tokens": 32768, "reference_input_tokens": 32768, "input_content_sha256": "25c9e190a2e46f598affb3fb2d368581a492c69f3d6dcb40b97ff719ef92b4bd", "render_sha256": "8741d3b84a76b641a93767e66d9a578272d32b01885611d215ce8cfcac71a546", "generated_tokens": 949, "generation_seconds_native": 18.154640674591064, "generation_tps_native": 52.27313594414483, "generation_tps_intertoken": 52.21805360911412, "input_seconds_native": 12.045980453491211, "input_tps_native": 2720.243497531416, "first_generated_seconds": 12.068147322977893, "first_visible_seconds": 12.068147322977893, "first_answer_visible_seconds": 13.875998331932351, "end_to_end_seconds": 30.207308548036963, "finish_reason": "stop_token", "cached_tokens": 0, "context_shift_or_requeue": false, "performance_integrity_checks_pass": true, "answer_quality_pass": false, "answer_quality_review": "FULL_OUTPUT_REVIEWED_BY_CODEX_SEPARATE_FROM_SPEED", "output_sha256": "517e4b39941582991e2f70eff357880c19725f0b5fc9dc597b9f310633de4072", "quality_findings": [ "記録00410の入力長を30,370としたが、実際は30,954。開始待ち0〜80秒も誤りで最小1秒。", "入力処理時間を送信から最初のトークンまでとしており、純粋なprefillを分離していない。" ] } ], "on": [ { "accepted_draft_tokens": 465, "rejected_draft_tokens": 510, "draft_acceptance_ratio": 0.47692307692307695, "case": "ctx8192-r1", "actual_input_tokens": 8193, "reference_input_tokens": 8192, "input_content_sha256": "fe7772257f2e32cecb8ed3052c1bc076e77e0c0b68e97fcb11cffb2f0035660c", "render_sha256": "2785e85f267cc7917263671a8b368ee917ff67ad3bc7e1a050abcfb98c4ab4dc", "generated_tokens": 790, "generation_seconds_native": 12.601438045501709, "generation_tps_native": 62.69125770784577, "generation_tps_intertoken": 62.61190168543078, "input_seconds_native": 3.824756383895874, "input_tps_native": 2142.09721552374, "first_generated_seconds": 4.515950244036503, "first_visible_seconds": 4.515950244036503, "first_answer_visible_seconds": 5.725612033973448, "end_to_end_seconds": 16.570731892017648, "finish_reason": "stop_token", "cached_tokens": 0, "context_shift_or_requeue": false, "performance_integrity_checks_pass": true, "answer_quality_pass": false, "answer_quality_review": "FULL_OUTPUT_REVIEWED_BY_CODEX_SEPARATE_FROM_SPEED", "output_sha256": "291bd85092dbc4d73f16cc3fa4bd172a12e7a3d46cd131f7b66922191e8a822d", "quality_findings": [ "GPUメモリ余裕を「コンテキスト長の20%未満」と比較しており、バイト数とトークン長の次元が合わない。", "入力処理時間を応答開始までの開始待ちで測っており、prefillとキュー・初回生成を分離していない。", "混雑原因は推測であると明示している点は保持する。" ] }, { "accepted_draft_tokens": 560, "rejected_draft_tokens": 538, "draft_acceptance_ratio": 0.5100182149362478, "case": "ctx32768-r1", "actual_input_tokens": 32768, "reference_input_tokens": 32768, "input_content_sha256": "5861ad9ee43affadb6a889494a29ef9097c485e5671850050a8e9f24c5214014", "render_sha256": "8741d3b84a76b641a93767e66d9a578272d32b01885611d215ce8cfcac71a546", "generated_tokens": 926, "generation_seconds_native": 12.31331729888916, "generation_tps_native": 75.20312987333955, "generation_tps_intertoken": 75.121916990107, "input_seconds_native": 13.206957578659058, "input_tps_native": 2481.1164724985083, "first_generated_seconds": 13.243030342971906, "first_visible_seconds": 13.243030342971906, "first_answer_visible_seconds": 13.888618264929391, "end_to_end_seconds": 25.52739283500705, "finish_reason": "stop_token", "cached_tokens": 0, "context_shift_or_requeue": false, "performance_integrity_checks_pass": true, "answer_quality_pass": false, "answer_quality_review": "FULL_OUTPUT_REVIEWED_BY_CODEX_SEPARATE_FROM_SPEED", "output_sha256": "5e7cf81b705691aa2837105edad625c6fb56f25f5df8523a0f41e4616de9e7b0", "quality_findings": [ "完全な記録00648もあるが、見出しは00647までとする。", "開始待ちを外部から1秒／80秒に設定する手順を提案しており、観測された待ち時間と負荷条件の区別が不十分。", "メモリ余裕20%以上を合格とする等の基準は本記録から検証されていない。" ] }, { "accepted_draft_tokens": 601, "rejected_draft_tokens": 587, "draft_acceptance_ratio": 0.5058922558922558, "case": "ctx32768-r2", "actual_input_tokens": 32768, "reference_input_tokens": 32768, "input_content_sha256": "df87c9d1b8895eaf28156c4236df58438ee505413bfa71c5071520e755fab152", "render_sha256": "8741d3b84a76b641a93767e66d9a578272d32b01885611d215ce8cfcac71a546", "generated_tokens": 997, "generation_seconds_native": 13.41328740119934, "generation_tps_native": 74.32928037543233, "generation_tps_intertoken": 74.25472743623932, "input_seconds_native": 12.403007745742798, "input_tps_native": 2641.9398158682334, "first_generated_seconds": 12.439681946998462, "first_visible_seconds": 12.439706002944149, "first_answer_visible_seconds": 12.439706002944149, "end_to_end_seconds": 25.82369013596326, "finish_reason": "stop_token", "cached_tokens": 0, "context_shift_or_requeue": false, "performance_integrity_checks_pass": true, "answer_quality_pass": false, "answer_quality_review": "FULL_OUTPUT_REVIEWED_BY_CODEX_SEPARATE_FROM_SPEED", "output_sha256": "5e611bf69e2d0d24e94713e0e8f1bc1f85d5ed47068f6676f65b10dc92dea92d", "quality_findings": [ "開始待ち0〜80秒は誤りで、最小値は1秒。", "生成速度の式が時間を出力トークン数で割る説明になっており、tokens/sの逆数。完了時刻とTTFTという期間も混在する。", "約1,000〜31,000を約1,000刻みで10段階とする試験水準の説明に不整合。", "入力処理時間の測定としてTTFTを使い、prefillを分離していない。" ] }, { "accepted_draft_tokens": 570, "rejected_draft_tokens": 660, "draft_acceptance_ratio": 0.4634146341463415, "case": "ctx8192-r2", "actual_input_tokens": 8193, "reference_input_tokens": 8192, "input_content_sha256": "227fa296fc7b644c484fc40877bdb35e587d70539a7bd314d37a43409cdf2dbb", "render_sha256": "2785e85f267cc7917263671a8b368ee917ff67ad3bc7e1a050abcfb98c4ab4dc", "generated_tokens": 980, "generation_seconds_native": 13.752604961395264, "generation_tps_native": 71.25922708831844, "generation_tps_intertoken": 71.18651359128955, "input_seconds_native": 3.116899251937866, "input_tps_native": 2628.5738927577386, "first_generated_seconds": 3.15280508401338, "first_visible_seconds": 3.1528311029542238, "first_answer_visible_seconds": 3.1528311029542238, "end_to_end_seconds": 16.875756316934712, "finish_reason": "stop_token", "cached_tokens": 0, "context_shift_or_requeue": false, "performance_integrity_checks_pass": true, "answer_quality_pass": false, "answer_quality_review": "FULL_OUTPUT_REVIEWED_BY_CODEX_SEPARATE_FROM_SPEED", "output_sha256": "f7296b28af09f33555f1f6b342a688b138d6ca46e036410931056e806e571b8b", "quality_findings": [ "開始待ち0〜80秒は誤りで最小1秒。4件同時の待ち2〜80秒も誤りで、実際は2〜78秒。", "prefillとTTFTを別手順で扱い、閾値を例と明示した点は保持する。" ] }, { "accepted_draft_tokens": 616, "rejected_draft_tokens": 605, "draft_acceptance_ratio": 0.5045045045045045, "case": "ctx8192-r3", "actual_input_tokens": 8193, "reference_input_tokens": 8192, "input_content_sha256": "e50d4d80b003558993083950a30517f1b8aa99cb95c542b50b33782237d011f6", "render_sha256": "2785e85f267cc7917263671a8b368ee917ff67ad3bc7e1a050abcfb98c4ab4dc", "generated_tokens": 1023, "generation_seconds_native": 13.556791305541992, "generation_tps_native": 75.4603340085201, "generation_tps_intertoken": 75.38657024116084, "input_seconds_native": 3.119197368621826, "input_tps_native": 2626.637250473176, "first_generated_seconds": 3.1550411159405485, "first_visible_seconds": 3.1550662230001763, "first_answer_visible_seconds": 3.1550662230001763, "end_to_end_seconds": 16.68231029296294, "finish_reason": "stop_token", "cached_tokens": 0, "context_shift_or_requeue": false, "performance_integrity_checks_pass": true, "answer_quality_pass": false, "answer_quality_review": "FULL_OUTPUT_REVIEWED_BY_CODEX_SEPARATE_FROM_SPEED", "output_sha256": "5b33c841c703f6e0e3ce339deea6aa6dd0501577c0c79f94b819357c15e20c72", "quality_findings": [ "入力処理時間をTTFTと同一視し、後段でも実行開始から最初のトークンまでをそのままprefillとしている。", "同時数増加で待ちが長くなる場合にGPUバッチ容量を原因と判断できるとしたが、他の要因を除外できない。", "一般のnvidia-smi使用量ログだけで断片化を直接測定する手順は示されていない。" ] }, { "accepted_draft_tokens": 526, "rejected_draft_tokens": 491, "draft_acceptance_ratio": 0.5172074729596854, "case": "ctx32768-r3", "actual_input_tokens": 32768, "reference_input_tokens": 32768, "input_content_sha256": "25c9e190a2e46f598affb3fb2d368581a492c69f3d6dcb40b97ff719ef92b4bd", "render_sha256": "8741d3b84a76b641a93767e66d9a578272d32b01885611d215ce8cfcac71a546", "generated_tokens": 865, "generation_seconds_native": 11.390297889709473, "generation_tps_native": 75.94182420650134, "generation_tps_intertoken": 75.85403019007764, "input_seconds_native": 12.410313367843628, "input_tps_native": 2640.3845760176523, "first_generated_seconds": 12.446683499962091, "first_visible_seconds": 12.446683499962091, "first_answer_visible_seconds": 13.65313025994692, "end_to_end_seconds": 23.808734862948768, "finish_reason": "stop_token", "cached_tokens": 0, "context_shift_or_requeue": false, "performance_integrity_checks_pass": true, "answer_quality_pass": false, "answer_quality_review": "FULL_OUTPUT_REVIEWED_BY_CODEX_SEPARATE_FROM_SPEED", "output_sha256": "e6cb00e89bca754c6aa833212702706093b64addac4d1b2c155cdffcd30b2cad", "quality_findings": [ "開始待ちが80秒で1秒に折り返すという説明は不正確。11秒加算の剰余で循環し、80秒の次は11秒。", "入力処理時間をTTFTで計測し、prefillを分離していない。", "待ち80秒・メモリ余裕10%・速度低下50%を不合格とする基準は実運用要件から検証されていない（末尾の架空データへの留保も保持）。" ] } ] }, "actual_peak_memory_MiB": { "off": { "0": 62947, "1": 62945 }, "on": { "0": 66789, "1": 62957 } }, "generation_cap_count": { "off": 0, "on": 0 }, "all_trial_averages_above50": { "off": true, "on": true }, "fixed_comparison_fields": [ "order", "context_capacity", "max_generation", "concurrency", "no_cpu_offload", "prefix_cache_reuse", "engine_variant", "model_revision", "transport", "actual_tokens_expected", "metric_definitions" ], "same_requests_and_rendered_prompts_as_B200": true, "sources": { "off": { "name": "exl3-main8k32k-independent-review.json", "sha256": "fc8165a3b8f83c81e004c76aeebfebf4cf8f0ba327689e07840ded0eb063fbf3" }, "on": { "name": "exl3-mtp3-main8k32k-independent-review.json", "sha256": "b46d657a381863f02add3e0bffb47df3fe7a8e7b5de66b801ff51e601b13a4cf" }, "quality": { "name": "all12-answer-quality-review.json", "sha256": "12dc2eebb0066a218c0387117b297ad91cdde9aec538173840307d10851ed46c" }, "H100_llamacpp": { "name": "llamacpp-mtp-final-comparison.json", "sha256": "25ff1866a56061c3cf7978ea61d6fc4a5b608bdd39cd47a797d99735166ab64d" } }, "metric_definitions": { "native_reported_tps": "new_tokens/time_generate", "conservative_intertoken_tps": "(new_tokens-1)/time_generate", "first_generated": "first streaming event with token IDs, including reasoning control tokens", "first_visible_any": "first nonempty text after removing markup; reasoning text counts", "first_answer_visible": "first nonempty visible text after closing think tag", "prefill": "native time_prefill includes input work up to first generation; prompt tokens/native time_prefill", "end_to_end": "enqueue to final event, excludes model load, warmup, disk writes and tokenizer" }, "limitations": [ "Three trials per condition and one synthetic workload; no instantaneous minimum, SLA or arbitrary-workload guarantee.", "All reasoning/control tokens count according to each engine; tokens/s is not visible Japanese characters per second.", "Off/on condition blocks are sequential, not randomized. First slow trial is retained; no separate untimed MTP generation before main timing.", "Each trial can end naturally at a different output length. End-to-end duration is not a fixed-output-length comparison.", "H100 uses NCCL_P2P_DISABLE=1 and shared host memory communication after default P2P stalled; B200 used its original transport. This is not an isolated GPU-architecture comparison.", "H100 PCIe PHB and B200 NVLink are different systems. GPU architecture, interconnect, base OS/Python patch release and power limits are not isolated causal effects.", "ExLlama EXL3 and llama.cpp GGUF UD-IQ3_XXS differ in quantization, tensor versus layer split and in-process versus HTTP interface.", "Temporary host model storage is distinct from inference offload; all actual weights and KV verified on GPUs. One 4-byte int32 format flag remains CPU metadata.", "Full responses are retained and separately reviewed; this small study cannot attribute quality differences to MTP alone. RTX PRO 6000 remains unmeasured." ] }