{"entries":[{"entry":{"configuration":"No speculation or stack patches","control_kind":"baseline","control_label":"Reference baseline","control_tps":null,"date":"13-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"stack-control","kind":"tps","lanes":32,"outcome":"","qualification":"Padded replay, not real task rollouts","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Measured","title":"Stack control","tps":476.8},"id":"stack-control","record":{"chart_comparable":false,"comparison":"Same-box stack baseline","configuration":"No speculation or stack patches","control_label":"Reference baseline","control_scope":"Baseline row","control_tps":null,"date":"13-Nov-2026","duration":"Not recorded per cell","id":"stack-control","improvement":"\u2014","improvement_pct":null,"kind":"tps","lanes":32,"outcome":"","qualification":"Padded replay, not real task rollouts","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Measured","title":"Stack control","tps":476.8},"version":"d8e5613710914990f5463a975267fe1111ee0947b0079fa856ce3bed8b22419d"},{"entry":{"configuration":"ngram_gpu k=3; thinking budget 4096 patch; sliding-window prefix-hit fix; tuned LoRA kernels","control_kind":"matched","control_label":"Stack control \u00b7 32 lanes","control_tps":476.8,"date":"13-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"stack32","kind":"tps","lanes":32,"outcome":"","qualification":"Aligned-engine C3 / lag-0 gates still required; stock decoding is not bitwise identical","recorded_window":"","run_url":"","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Measured","summary":"","title":"Combined stack","tps":716.0},"id":"stack32","record":{"chart_comparable":true,"comparison":"Same-study recorded control","configuration":"ngram_gpu k=3; thinking budget 4096 patch; sliding-window prefix-hit fix; tuned LoRA kernels","control_kind":"matched","control_label":"Stack control \u00b7 32 lanes","control_scope":"Same-study recorded control","control_tps":476.8,"date":"13-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"stack32","improvement":"+50.2%","improvement_pct":50.1678,"kind":"tps","lanes":32,"outcome":"","qualification":"Aligned-engine C3 / lag-0 gates still required; stock decoding is not bitwise identical","recorded_window":"","run_url":"","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","stack":"ngram_gpu k=3; thinking budget 4096 patch; sliding-window prefix-hit fix; tuned LoRA kernels","status":"Measured","summary":"","title":"Combined stack","tps":716.0},"version":"f1221011727bf6c764373fa9adbad2dcae9f4a6108e0c8aa0643319599978315"},{"entry":{"configuration":"Same stack; async scheduling off","control_kind":"matched","control_label":"Stack control \u00b7 32 lanes","control_tps":476.8,"date":"13-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"stack48","kind":"tps","lanes":48,"outcome":"","qualification":"KV peak 99.9%; replay success did not transfer to full runs","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Measured","title":"Combined stack, async off","tps":735.5},"id":"stack48","record":{"chart_comparable":true,"comparison":"+54.3% vs 476.8 same-box control","configuration":"Same stack; async scheduling off","control_label":"Stack control \u00b7 32 lanes","control_scope":"Same box; lane count differs where stated","control_tps":476.8,"date":"13-Nov-2026","duration":"Not recorded per cell","id":"stack48","improvement":"+54.3%","improvement_pct":54.3,"kind":"tps","lanes":48,"outcome":"","qualification":"KV peak 99.9%; replay success did not transfer to full runs","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Measured","title":"Combined stack, async off","tps":735.5},"version":"724ba8a944a04bf3a2cfb9aba0de66e670be09fe0d88edaa60f2a9b4c6df8ccb"},{"entry":{"configuration":"Same stack; fp8 KV cache","control_kind":"matched","control_label":"Stack control \u00b7 32 lanes","control_tps":476.8,"date":"13-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"fp8-32","kind":"tps","lanes":32,"outcome":"","qualification":"Lossy KV changes numerics; measurement only, not approved for training","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Measured","title":"Combined stack + fp8 KV","tps":849.6},"id":"fp8-32","record":{"chart_comparable":true,"comparison":"+78.2% vs 476.8 same-box control","configuration":"Same stack; fp8 KV cache","control_label":"Stack control \u00b7 32 lanes","control_scope":"Same box; lane count differs where stated","control_tps":476.8,"date":"13-Nov-2026","duration":"Not recorded per cell","id":"fp8-32","improvement":"+78.2%","improvement_pct":78.2,"kind":"tps","lanes":32,"outcome":"","qualification":"Lossy KV changes numerics; measurement only, not approved for training","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Measured","title":"Combined stack + fp8 KV","tps":849.6},"version":"536c80df81db1e2c86bf737686c30a87f5cb304b1c8c6e2f49a89dc7aca579be"},{"entry":{"configuration":"Same stack; fp8 KV; async scheduling off","control_kind":"matched","control_label":"Stack control \u00b7 32 lanes","control_tps":476.8,"date":"13-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"fp8-48","kind":"tps","lanes":48,"outcome":"","qualification":"Lossy KV; KV peak 50%; not a real-rollout result","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Measured","title":"Combined stack + fp8 KV, async off","tps":965.7},"id":"fp8-48","record":{"chart_comparable":true,"comparison":"+102.5% vs 476.8 same-box control","configuration":"Same stack; fp8 KV; async scheduling off","control_label":"Stack control \u00b7 32 lanes","control_scope":"Same box; lane count differs where stated","control_tps":476.8,"date":"13-Nov-2026","duration":"Not recorded per cell","id":"fp8-48","improvement":"+102.5%","improvement_pct":102.5,"kind":"tps","lanes":48,"outcome":"","qualification":"Lossy KV; KV peak 50%; not a real-rollout result","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Measured","title":"Combined stack + fp8 KV, async off","tps":965.7},"version":"8ffed5fd4d7e3dddfdaf17292089a7aaae95391f9c5ef9a5f68082bf13f945a8"},{"entry":{"configuration":"ngram_gpu k=3; budget 4096 patch","control_kind":"matched","control_label":"Own control \u00b7 budget 4096","control_tps":479.7,"date":"13-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"ngram-budget","kind":"tps","lanes":32,"outcome":"","qualification":"Same distribution in principle; aligned-engine checks needed","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#results","status":"Measured","title":"ngram + preserved thinking budget","tps":580.5},"id":"ngram-budget","record":{"chart_comparable":true,"comparison":"+21.0% vs own control (479.7)","configuration":"ngram_gpu k=3; budget 4096 patch","control_label":"Own control \u00b7 budget 4096","control_scope":"Same-study control","control_tps":479.7,"date":"13-Nov-2026","duration":"Not recorded per cell","id":"ngram-budget","improvement":"+21.0%","improvement_pct":21.0,"kind":"tps","lanes":32,"outcome":"","qualification":"Same distribution in principle; aligned-engine checks needed","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#results","status":"Measured","title":"ngram + preserved thinking budget","tps":580.5},"version":"9ebb229536af2bc4ff62d2e02aa63d8e294d7acc23cb5937e7fb2b6382dea9d4"},{"entry":{"configuration":"ngram_gpu k=3; sliding-window prefix-hit fix","control_kind":"unknown","control_label":"Not recorded","control_tps":null,"date":"13-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"prefix-ngram","kind":"tps","lanes":32,"outcome":"","qualification":"Prefix hits are not bitwise recomputation on stock vLLM; C3 needed","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#results","status":"Measured","title":"Prefix-hit fix + ngram","tps":701.3},"id":"prefix-ngram","record":{"chart_comparable":false,"comparison":"Source reports a relative gain, but exact control TPS is not recorded in this summary; no numeric improvement shown.","configuration":"ngram_gpu k=3; sliding-window prefix-hit fix","control_label":"Not recorded","control_scope":"No numeric improvement claimed","control_tps":null,"date":"13-Nov-2026","duration":"Not recorded per cell","id":"prefix-ngram","improvement":"\u2014","improvement_pct":null,"kind":"tps","lanes":32,"outcome":"","qualification":"Prefix hits are not bitwise recomputation on stock vLLM; C3 needed","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#results","status":"Measured","title":"Prefix-hit fix + ngram","tps":701.3},"version":"360d69645369bb82ea4662a8cda59e136d98eb2313be91f23500945dfff36403"},{"entry":{"configuration":"Combined stack; latest-match ngram and adaptive draft length","control_kind":"matched","control_label":"Combined stack before latest-match change","control_tps":713.7,"date":"13-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"latest","kind":"tps","lanes":32,"outcome":"","qualification":"Adaptive length alone within noise; replay only","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#since-the-big-run-other-sessions-all-measured-on-keith-exact-or-replay-setups","status":"Measured","title":"Latest-match drafting + adaptive length","tps":756.4},"id":"latest","record":{"chart_comparable":true,"comparison":"+6.0% vs 713.7 in two matched pairs","configuration":"Combined stack; latest-match ngram and adaptive draft length","control_label":"Combined stack before latest-match change","control_scope":"Two matched replay pairs","control_tps":713.7,"date":"13-Nov-2026","duration":"Not recorded per cell","id":"latest","improvement":"+6.0%","improvement_pct":6.0,"kind":"tps","lanes":32,"outcome":"","qualification":"Adaptive length alone within noise; replay only","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#since-the-big-run-other-sessions-all-measured-on-keith-exact-or-replay-setups","status":"Measured","title":"Latest-match drafting + adaptive length","tps":756.4},"version":"bc1c3cc6d16d7dc61fe3670527806a9619934edcec81f6cd71fb552d6d77ad59"},{"entry":{"configuration":"ngram async fix; async scheduling on","control_kind":"unknown","control_label":"Not recorded","control_tps":null,"date":"13-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"async-fix","kind":"tps","lanes":48,"outcome":"","qualification":"Acceptance restored to 1.90; does not remove real-rollout KV limit","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#since-the-big-run-other-sessions-all-measured-on-keith-exact-or-replay-setups","status":"Measured","title":"Async staging-buffer fix","tps":760.1},"id":"async-fix","record":{"chart_comparable":false,"comparison":"Source compares async on with async off, but exact async-off control TPS is not recorded in this summary; no numeric improvement shown.","configuration":"ngram async fix; async scheduling on","control_label":"Not recorded","control_scope":"No numeric improvement claimed","control_tps":null,"date":"13-Nov-2026","duration":"Not recorded per cell","id":"async-fix","improvement":"\u2014","improvement_pct":null,"kind":"tps","lanes":48,"outcome":"","qualification":"Acceptance restored to 1.90; does not remove real-rollout KV limit","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#since-the-big-run-other-sessions-all-measured-on-keith-exact-or-replay-setups","status":"Measured","title":"Async staging-buffer fix","tps":760.1},"version":"ac4c1d432dde5cad2b2fc5574094a7ba61816127857178e50f2ee8eddf2f585b"},{"entry":{"configuration":"fp8; k=3; async off; CUDA graphs to 1024; four turns","control_kind":"unknown","control_label":"Not recorded","control_tps":null,"date":"13-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"fp8-64","kind":"tps","lanes":64,"outcome":"","qualification":"Shorter prompts than the main bench; lossy KV","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Measured","title":"fp8 scaling, shorter replay","tps":1075.5},"id":"fp8-64","record":{"chart_comparable":false,"comparison":"No numeric improvement claimed","configuration":"fp8; k=3; async off; CUDA graphs to 1024; four turns","control_label":"Not recorded","control_scope":"No numeric improvement claimed","control_tps":null,"date":"13-Nov-2026","duration":"Not recorded per cell","id":"fp8-64","improvement":"\u2014","improvement_pct":null,"kind":"tps","lanes":64,"outcome":"","qualification":"Shorter prompts than the main bench; lossy KV","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Measured","title":"fp8 scaling, shorter replay","tps":1075.5},"version":"5c4bbe81e72d7ef1741ca362cff61b2a428b51098c7a32d5c10dc0c1d1dd30ad"},{"entry":{"configuration":"Same four-turn setup","control_kind":"matched","control_label":"64 lanes \u00b7 same four-turn replay","control_tps":1075.5,"date":"13-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"fp8-96","kind":"tps","lanes":96,"outcome":"","qualification":"Per-lane decode falls to 10.5; lossy KV","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Measured","title":"fp8 scaling, shorter replay","tps":1093.9},"id":"fp8-96","record":{"chart_comparable":true,"comparison":"Near plateau vs 1075.5 at 64 lanes","configuration":"Same four-turn setup","control_label":"64 lanes \u00b7 same four-turn replay","control_scope":"Calculated from recorded TPS; different concurrency","control_tps":1075.5,"date":"13-Nov-2026","duration":"Not recorded per cell","id":"fp8-96","improvement":"+1.7%","improvement_pct":1.7,"kind":"tps","lanes":96,"outcome":"","qualification":"Per-lane decode falls to 10.5; lossy KV","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Measured","title":"fp8 scaling, shorter replay","tps":1093.9},"version":"256291f6b4a5de6d4de4794074909da99cf13468e6edd75d7d50e02d8ef448c2"},{"entry":{"configuration":"vLLM 0.32; no speculation","control_kind":"historical","control_label":"Earlier v0.19.1 baseline (approx.)","control_tps":476,"date":"13-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"v032-base","kind":"tps","lanes":32,"outcome":"","qualification":"Version comparison, not an aligned-engine gate","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#vllm-032-and-4-bit-kv-r31b-kv4-1-own-venv-lossy-kv-types-are-measurements-only","status":"Measured","title":"vLLM 0.32 baseline","tps":500.9},"id":"v032-base","record":{"chart_comparable":false,"comparison":"0.19.1 baseline approximately 476; different engine","configuration":"vLLM 0.32; no speculation","control_label":"Earlier v0.19.1 baseline (approx.)","control_scope":"Historical engine comparison; approximate TPS","control_tps":476,"date":"13-Nov-2026","duration":"Not recorded per cell","id":"v032-base","improvement":"Not matched","improvement_pct":null,"kind":"tps","lanes":32,"outcome":"","qualification":"Version comparison, not an aligned-engine gate","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#vllm-032-and-4-bit-kv-r31b-kv4-1-own-venv-lossy-kv-types-are-measurements-only","status":"Measured","title":"vLLM 0.32 baseline","tps":500.9},"version":"8d0e9d8498a476e8504613fbefc8f2984912dedbbf950e69c2a53f27205fa7c9"},{"entry":{"configuration":"vLLM 0.32; ngram k=3; bf16 KV","control_kind":"matched","control_label":"v0.32 \u00b7 no speculation","control_tps":500.9,"date":"13-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"v032-ngram","kind":"tps","lanes":32,"outcome":"","qualification":"Replay only","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#vllm-032-and-4-bit-kv-r31b-kv4-1-own-venv-lossy-kv-types-are-measurements-only","status":"Measured","title":"vLLM 0.32 ngram","tps":599.0},"id":"v032-ngram","record":{"chart_comparable":true,"comparison":"Calculated from same-study recorded TPS","configuration":"vLLM 0.32; ngram k=3; bf16 KV","control_label":"v0.32 \u00b7 no speculation","control_scope":"Calculated from same-study recorded TPS","control_tps":500.9,"date":"13-Nov-2026","duration":"Not recorded per cell","id":"v032-ngram","improvement":"+19.6%","improvement_pct":19.6,"kind":"tps","lanes":32,"outcome":"","qualification":"Replay only","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#vllm-032-and-4-bit-kv-r31b-kv4-1-own-venv-lossy-kv-types-are-measurements-only","status":"Measured","title":"vLLM 0.32 ngram","tps":599.0},"version":"b6f209aa5a4d22a45e25aaf6af47c878f48128b5f8a3ce84d175434c97618e16"},{"entry":{"configuration":"vLLM 0.32; ngram k=3; int4_per_token_head; Triton","control_kind":"matched","control_label":"v0.32 \u00b7 bf16 KV + ngram","control_tps":599.0,"date":"13-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"v032-int4","kind":"tps","lanes":32,"outcome":"","qualification":"Lossy KV; unpacking costs outweigh smaller reads","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#vllm-032-and-4-bit-kv-r31b-kv4-1-own-venv-lossy-kv-types-are-measurements-only","status":"Measured","title":"vLLM 0.32 int4 KV","tps":560.7},"id":"v032-int4","record":{"chart_comparable":true,"comparison":"Calculated from same-study recorded TPS","configuration":"vLLM 0.32; ngram k=3; int4_per_token_head; Triton","control_label":"v0.32 \u00b7 bf16 KV + ngram","control_scope":"Calculated from same-study recorded TPS","control_tps":599.0,"date":"13-Nov-2026","duration":"Not recorded per cell","id":"v032-int4","improvement":"\u22126.4%","improvement_pct":-6.4,"kind":"tps","lanes":32,"outcome":"","qualification":"Lossy KV; unpacking costs outweigh smaller reads","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#vllm-032-and-4-bit-kv-r31b-kv4-1-own-venv-lossy-kv-types-are-measurements-only","status":"Measured","title":"vLLM 0.32 int4 KV","tps":560.7},"version":"d4543d08b0411ee5097ce525009d61ea1f60c5f2eaab658db02e52032161bf94"},{"entry":{"configuration":"vLLM 0.32; FlashInfer; fp8; ngram k=3; r16; budget off","control_kind":"matched","control_label":"FlashInfer bf16 \u00b7 same box","control_tps":617.0,"date":"14-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"fi-fp8","kind":"tps","lanes":32,"outcome":"","qualification":"Lossy KV; different engine; budget patch absent","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#nvfp4-kv-on-flashinfer-r31b-kv4e-7-14-nov","status":"Measured","title":"FlashInfer fp8 KV","tps":811.7},"id":"fi-fp8","record":{"chart_comparable":true,"comparison":"Calculated from same-backend recorded TPS; lossy KV","configuration":"vLLM 0.32; FlashInfer; fp8; ngram k=3; r16; budget off","control_label":"FlashInfer bf16 \u00b7 same box","control_scope":"Calculated from same-backend recorded TPS; lossy KV","control_tps":617.0,"date":"14-Nov-2026","duration":"Not recorded per cell","id":"fi-fp8","improvement":"+31.6%","improvement_pct":31.6,"kind":"tps","lanes":32,"outcome":"","qualification":"Lossy KV; different engine; budget patch absent","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#nvfp4-kv-on-flashinfer-r31b-kv4e-7-14-nov","status":"Measured","title":"FlashInfer fp8 KV","tps":811.7},"version":"bff46e82289ed32546d6f4ba9f16b4fa4202837dcc87fc16521033fdc1061131"},{"entry":{"configuration":"vLLM 0.32; text-only multimodal-prefix patch; nvfp4; ngram k=3; r16; budget off","control_kind":"matched","control_label":"FlashInfer fp8 \u00b7 same box","control_tps":811.7,"date":"14-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"fi-fp4","kind":"tps","lanes":32,"outcome":"","qualification":"Lossy 4-bit KV; numerical drift unverified; more room is not proven end-to-end speed","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#nvfp4-kv-on-flashinfer-r31b-kv4e-7-14-nov","status":"Measured","title":"FlashInfer nvfp4 KV","tps":699.6},"id":"fi-fp4","record":{"chart_comparable":true,"comparison":"14% slower than fp8 on the same backend; 1.78\u00d7 KV capacity","configuration":"vLLM 0.32; text-only multimodal-prefix patch; nvfp4; ngram k=3; r16; budget off","control_label":"FlashInfer fp8 \u00b7 same box","control_scope":"Source-reported same-backend comparison; lossy KV","control_tps":811.7,"date":"14-Nov-2026","duration":"Not recorded per cell","id":"fi-fp4","improvement":"\u221214%","improvement_pct":-14.0,"kind":"tps","lanes":32,"outcome":"","qualification":"Lossy 4-bit KV; numerical drift unverified; more room is not proven end-to-end speed","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#nvfp4-kv-on-flashinfer-r31b-kv4e-7-14-nov","status":"Measured","title":"FlashInfer nvfp4 KV","tps":699.6},"version":"89e5aea11344124129bfd7f3bda508b5da8eb9a8efabc4174be49fae81cb432a"},{"entry":{"configuration":"Combined stack; fp8 KV; async off","control_kind":"unknown","control_label":"Not recorded","control_tps":null,"date":"13-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"fp8-64-interrupted","kind":"tps","lanes":64,"outcome":"","qualification":"Box preempted; cancelled to free GPU","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Interrupted","title":"64-lane combined-stack fp8 attempt","tps":null},"id":"fp8-64-interrupted","record":{"chart_comparable":false,"comparison":"No completed measurement","configuration":"Combined stack; fp8 KV; async off","control_label":"Not recorded","control_scope":"No numeric improvement claimed","control_tps":null,"date":"13-Nov-2026","duration":"Not recorded per cell","id":"fp8-64-interrupted","improvement":"\u2014","improvement_pct":null,"kind":"tps","lanes":64,"outcome":"","qualification":"Box preempted; cancelled to free GPU","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Interrupted","title":"64-lane combined-stack fp8 attempt","tps":null},"version":"0a8707ec197ebc924daf68e01f25f025cd6b132c264161b94b177a828ac6103b"},{"entry":{"configuration":"Four-turn fp8 setup; k=1 / k=2 / k=3","control_kind":"unknown","control_label":"Not recorded","control_tps":null,"date":"13-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"lanes128","kind":"tps","lanes":128,"outcome":"","qualification":"k=1/2 shape-mismatch crash; k=3 CUDA out of memory","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Failed","title":"128-lane scaling attempts","tps":null},"id":"lanes128","record":{"chart_comparable":false,"comparison":"No throughput result","configuration":"Four-turn fp8 setup; k=1 / k=2 / k=3","control_label":"Not recorded","control_scope":"No numeric improvement claimed","control_tps":null,"date":"13-Nov-2026","duration":"Not recorded per cell","id":"lanes128","improvement":"\u2014","improvement_pct":null,"kind":"tps","lanes":128,"outcome":"","qualification":"k=1/2 shape-mismatch crash; k=3 CUDA out of memory","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Failed","title":"128-lane scaling attempts","tps":null},"version":"81dc66cb53f795d31634dccc4ddb2171521731e75ba119b9fc9e2d17b75d71eb"},{"entry":{"configuration":"No speculation; no adapter; budget_aware_v2_plus_31b; temperature 1; 7.5-minute cap; eight dev-v1 tasks \u00d7 eight rollouts","control_kind":"baseline","control_label":"Reference baseline \u00b7 arm A","control_tps":null,"date":"13-Nov-2026","duration":"895 s server-busy measurement window; not whole-job wall time","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"real-a","kind":"full","lanes":32,"outcome":"36/64 solved (56%); 83% hit cap; 251.7 rollouts/hour","qualification":"Real Docker sandboxes and grading. Not a golden evaluation or promotion.","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#real-rollouts-one-box-64-rollouts-per-arm-rollout31b-real-1-with-the-container-session","status":"Completed","summary":"Complete \u00b7 36/64 solved","title":"Real-task baseline \u00b7 arm A","tps":384.8},"id":"real-a","record":{"chart_comparable":false,"comparison":"Baseline for arm B","configuration":"No speculation; no adapter; budget_aware_v2_plus_31b; temperature 1; 7.5-minute cap; eight dev-v1 tasks \u00d7 eight rollouts","control_label":"Reference baseline \u00b7 arm A","control_scope":"Baseline row","control_tps":null,"date":"13-Nov-2026","duration":"895 s server-busy measurement window; not whole-job wall time","id":"real-a","improvement":"\u2014","improvement_pct":null,"kind":"full","lanes":32,"outcome":"36/64 solved (56%); 83% hit cap; 251.7 rollouts/hour","qualification":"Real Docker sandboxes and grading. Not a golden evaluation or promotion.","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#real-rollouts-one-box-64-rollouts-per-arm-rollout31b-real-1-with-the-container-session","status":"Completed \u00b7 64 rollouts","summary":"Complete \u00b7 36/64 solved","title":"Real-task baseline \u00b7 arm A","tps":384.8},"version":"26c068ba8f416290057cc0715d88b3a2abf8dcab83a37bc5742c144636faaadb"},{"entry":{"configuration":"ngram k=3 + thinking budget patch; no adapter; budget_aware_v2_plus_31b; temperature 1; 7.5-minute cap; eight dev-v1 tasks \u00d7 eight rollouts","control_kind":"matched","control_label":"Real-task arm A \u00b7 stock","control_tps":384.8,"date":"13-Nov-2026","duration":"895 s server-busy measurement window; not whole-job wall time","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"real-b","kind":"full","lanes":32,"outcome":"33/64 solved (52%); 77% hit cap; 251.7 rollouts/hour","qualification":"Real Docker sandboxes and grading. Not a golden evaluation or promotion.","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#real-rollouts-one-box-64-rollouts-per-arm-rollout31b-real-1-with-the-container-session","status":"Completed","summary":"Complete \u00b7 33/64 solved","title":"Real-task ngram + budget \u00b7 arm B","tps":431.2},"id":"real-b","record":{"chart_comparable":true,"comparison":"+12.1% server-busy TPS vs arm A; solve difference within noise","configuration":"ngram k=3 + thinking budget patch; no adapter; budget_aware_v2_plus_31b; temperature 1; 7.5-minute cap; eight dev-v1 tasks \u00d7 eight rollouts","control_label":"Real-task arm A \u00b7 stock","control_scope":"Same-box paired task arms","control_tps":384.8,"date":"13-Nov-2026","duration":"895 s server-busy measurement window; not whole-job wall time","id":"real-b","improvement":"+12.1%","improvement_pct":12.1,"kind":"full","lanes":32,"outcome":"33/64 solved (52%); 77% hit cap; 251.7 rollouts/hour","qualification":"Real Docker sandboxes and grading. Not a golden evaluation or promotion.","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#real-rollouts-one-box-64-rollouts-per-arm-rollout31b-real-1-with-the-container-session","status":"Completed \u00b7 64 rollouts","summary":"Complete \u00b7 33/64 solved","title":"Real-task ngram + budget \u00b7 arm B","tps":431.2},"version":"2705a124c0e63b6476db1972363f39934da17e352433554c91b98ed3f7a29e0a"},{"entry":{"configuration":"Combined stack; zero-output r16 policy LoRA; 60 dev-v1 tasks \u00d7 eight rollouts planned","control_kind":"historical","control_label":"Earlier stock arm A \u00b7 64 rollouts","control_tps":384.8,"date":"13-Nov-2026","duration":"Not recorded per cell","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"real-207","kind":"full","lanes":32,"outcome":"58.5% resolved in this partial attempt; roughly 253 rollouts/hour","qualification":"Partial attempt; do not mix with restarted complete run","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Interrupted","summary":"Preempted \u00b7 207/480 \u00b7 58.5% resolved","title":"Combined stack \u00b7 preempted first attempt","tps":496.1},"id":"real-207","record":{"chart_comparable":false,"comparison":"+29% vs stock 384.8 in the reported table","configuration":"Combined stack; zero-output r16 policy LoRA; 60 dev-v1 tasks \u00d7 eight rollouts planned","control_label":"Earlier stock arm A \u00b7 64 rollouts","control_scope":"Historical / not matched: partial 207-rollout attempt vs earlier 64-rollout arm","control_tps":384.8,"date":"13-Nov-2026","duration":"Not recorded per cell","id":"real-207","improvement":"+29% reported","improvement_pct":null,"kind":"full","lanes":32,"outcome":"58.5% resolved in this partial attempt; roughly 253 rollouts/hour","qualification":"Partial attempt; do not mix with restarted complete run","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#everything-stacked-in-one-run-r31b-big-1","status":"Preempted \u00b7 207/480 rollouts","summary":"Preempted \u00b7 207/480 \u00b7 58.5% resolved","title":"Combined stack \u00b7 preempted first attempt","tps":496.1},"version":"9edfc7f230d4203306682ae48825db50ac125ae4649224a4c0b4626f53518f70"},{"entry":{"configuration":"Combined stack; zero-output r16 policy LoRA; 60 dev-v1 tasks \u00d7 eight; temperature 1; 7.5-minute cap","control_kind":"historical","control_label":"Earlier stock arm A \u00b7 64 rollouts","control_tps":384.8,"date":"13-Nov-2026","duration":"108.5 min whole rollout attempt; first 256-rollout update ready at 61.1 min","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"real-480","kind":"full","lanes":32,"outcome":"260/479 solved (54%); 265 rollouts/hour overall, 254/hour steady; 70% capped; KV peak 91%; no preemptions","qualification":"479 graded outcomes, one infra drop. Zero-output LoRA pays adapter compute but is not a trained policy. Aligned-engine compatibility remains unmeasured. Linked saved-run browser reports 261/480 solved, while the pinned throughput report states 260/479 after excluding one infra drop. Counts differ and have not been reconciled; retain both rather than treat them as identical scoring.","recorded_window":"Saved run metadata: 13-Nov 22:19 to 14-Nov 00:07, America/New_York (EST)","run_url":"/runs/4d62db5e2122f59fca76ed4b1842e9e3b379fc01f1f3727ad688dc1c549d3a5d","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#for-keith-480-real-rollouts-on-one-rtx-pro-6000-with-the-exact-stack-rollout31b-480-1","status":"Completed","summary":"Complete \u00b7 report 260/479; browser 261/480","title":"Combined stack \u00b7 complete 480-rollout restart","tps":485.3},"id":"real-480","record":{"chart_comparable":false,"comparison":"+26% server-busy TPS vs stock 384.8; cap still limits throughput","configuration":"Combined stack; zero-output r16 policy LoRA; 60 dev-v1 tasks \u00d7 eight; temperature 1; 7.5-minute cap","control_label":"Earlier stock arm A \u00b7 64 rollouts","control_scope":"Historical / not matched: complete 480-rollout run vs earlier 64-rollout arm","control_tps":384.8,"date":"13-Nov-2026","duration":"108.5 min whole rollout attempt; first 256-rollout update ready at 61.1 min","id":"real-480","improvement":"+26% reported","improvement_pct":null,"kind":"full","lanes":32,"outcome":"260/479 solved (54%); 265 rollouts/hour overall, 254/hour steady; 70% capped; KV peak 91%; no preemptions","qualification":"479 graded outcomes, one infra drop. Zero-output LoRA pays adapter compute but is not a trained policy. Aligned-engine compatibility remains unmeasured. Linked saved-run browser reports 261/480 solved, while the pinned throughput report states 260/479 after excluding one infra drop. Counts differ and have not been reconciled; retain both rather than treat them as identical scoring.","recorded_window":"Saved run metadata: 13-Nov 22:19 to 14-Nov 00:07, America/New_York (EST)","run_url":"/runs/4d62db5e2122f59fca76ed4b1842e9e3b379fc01f1f3727ad688dc1c549d3a5d","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#for-keith-480-real-rollouts-on-one-rtx-pro-6000-with-the-exact-stack-rollout31b-480-1","status":"Completed \u00b7 480 rollouts","summary":"Complete \u00b7 report 260/479; browser 261/480","title":"Combined stack \u00b7 complete 480-rollout restart","tps":485.3},"version":"a66aa04fafa38b51d73f9eae4cbb8f16e7c20b5ef5188c2b53bb169e3b364ebc"},{"entry":{"configuration":"Same 480 planned rollouts; combined stack + async fix + latest-match drafting; bf16 KV","control_kind":"unmatched","control_label":"32-lane complete run","control_tps":485.3,"date":"14-Nov-2026","duration":"51 min before cancellation; 36 groups of eight complete","hardware":"One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement","id":"real-48","kind":"full","lanes":48,"outcome":"99% capped; 82% empty patches; KV 100%; 173 preemptions; prefix hits 49.7%; 367 rollouts/hour over first 2838 s","qualification":"Do not compare whole-window server metrics as if time-matched. Faster rollouts/hour here means capped, truncated trajectories, not better useful output.","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#for-keith-48-lanes-on-real-rollouts-is-worse-than-32-rollout31b-480x48-1-14-nov","status":"Cancelled","summary":"Cancelled \u00b7 289/480 \u00b7 10.7% solved","title":"48 lanes \u00b7 cancelled real-task attempt","tps":281.7},"id":"real-48","record":{"chart_comparable":false,"comparison":"Server-busy TPS 281.7 vs 485.3 over each run\u2019s full busy window. Same 289 task/rollout keys: 10.7% solved vs 52.6% at 32 lanes.","configuration":"Same 480 planned rollouts; combined stack + async fix + latest-match drafting; bf16 KV","control_label":"32-lane complete run","control_scope":"Different whole-run busy windows; no TPS improvement inferred","control_tps":485.3,"date":"14-Nov-2026","duration":"51 min before cancellation; 36 groups of eight complete","id":"real-48","improvement":"Not matched","improvement_pct":null,"kind":"full","lanes":48,"outcome":"99% capped; 82% empty patches; KV 100%; 173 preemptions; prefix hits 49.7%; 367 rollouts/hour over first 2838 s","qualification":"Do not compare whole-window server metrics as if time-matched. Faster rollouts/hour here means capped, truncated trajectories, not better useful output.","source":"https://github.com/Autoresearch-Arena/gemma-4-developer-agent/blob/f38247e6aee9007e2dc41fbf67581759759465a8/reports/2026-11-13-31b-rollout-throughput-round2.md#for-keith-48-lanes-on-real-rollouts-is-worse-than-32-rollout31b-480x48-1-14-nov","status":"Cancelled \u00b7 289/480 rollouts","summary":"Cancelled \u00b7 289/480 \u00b7 10.7% solved","title":"48 lanes \u00b7 cancelled real-task attempt","tps":281.7},"version":"6282656b432db76e700cb16503f26aaa3e71e6720e89596724b6eae52bdd234a"}],"schema":"gemma-throughput-feed-v1"}