{ "concurrency": { "1": { "aggregate_end_to_end_tokens_per_second": { "maximum": 52.386, "median": 51.326, "minimum": 49.405, "p90": 52.386 }, "per_stream_decode_tokens_per_second": { "maximum": 54.733, "median": 53.624, "minimum": 51.444, "p90": 54.733 }, "per_stream_ttft_seconds": { "maximum": 0.232, "median": 0.227, "minimum": 0.224, "p90": 0.232 }, "rounds": [ { "aggregate_end_to_end_tokens_per_second": 52.386, "completion_tokens": 256, "round": 1, "streams": [ { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 4.658974, "decode_tokens_per_second": 54.733, "finish_reason": "length", "measured_sse_events": 114, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 181, "reasoning_characters": 1150, "reasoning_sha256": "d605239a7abf9e6823b6adc41abcee49b9aefd34629d57483888de5ee54f55a4", "response_tail_seconds": 0.000104, "stream_done_marker": true, "stream_sha256": "d605239a7abf9e6823b6adc41abcee49b9aefd34629d57483888de5ee54f55a4", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 4.886059, "ttft_seconds": 0.227085, "wall_seconds": 4.886163 } ], "wall_seconds": 4.886844 }, { "aggregate_end_to_end_tokens_per_second": 49.405, "completion_tokens": 256, "round": 2, "streams": [ { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 4.95687, "decode_tokens_per_second": 51.444, "finish_reason": "length", "measured_sse_events": 120, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 180, "reasoning_characters": 1095, "reasoning_sha256": "d3ba3995b56e19dfb3a22f5fa151b769f97028b54c0131c2a78d355ded71970d", "response_tail_seconds": 9.7e-05, "stream_done_marker": true, "stream_sha256": "d3ba3995b56e19dfb3a22f5fa151b769f97028b54c0131c2a78d355ded71970d", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 5.180915, "ttft_seconds": 0.224045, "wall_seconds": 5.181011 } ], "wall_seconds": 5.181659 }, { "aggregate_end_to_end_tokens_per_second": 51.326, "completion_tokens": 256, "round": 3, "streams": [ { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 4.755352, "decode_tokens_per_second": 53.624, "finish_reason": "length", "measured_sse_events": 117, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 181, "reasoning_characters": 1130, "reasoning_sha256": "b9df45e5056f6f8f8f00593e2f11dc4ba922fc59d37cda90c87d2fd3f783cf49", "response_tail_seconds": 0.000158, "stream_done_marker": true, "stream_sha256": "b9df45e5056f6f8f8f00593e2f11dc4ba922fc59d37cda90c87d2fd3f783cf49", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 4.987022, "ttft_seconds": 0.23167, "wall_seconds": 4.987181 } ], "wall_seconds": 4.987691 } ] }, "2": { "aggregate_end_to_end_tokens_per_second": { "maximum": 86.28, "median": 80.753, "minimum": 77.184, "p90": 86.28 }, "per_stream_decode_tokens_per_second": { "maximum": 47.016, "median": 43.632, "minimum": 41.526, "p90": 45.774 }, "per_stream_ttft_seconds": { "maximum": 0.512, "median": 0.361, "minimum": 0.214, "p90": 0.509 }, "rounds": [ { "aggregate_end_to_end_tokens_per_second": 80.753, "completion_tokens": 512, "round": 1, "streams": [ { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 5.827012, "decode_tokens_per_second": 43.762, "finish_reason": "length", "measured_sse_events": 117, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 178, "reasoning_characters": 1130, "reasoning_sha256": "3436805bf4e664884d74e8e48101324a57d621f6f729ac3dc203ceb85bfe57a9", "response_tail_seconds": 8.1e-05, "stream_done_marker": true, "stream_sha256": "3436805bf4e664884d74e8e48101324a57d621f6f729ac3dc203ceb85bfe57a9", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 6.338983, "ttft_seconds": 0.511971, "wall_seconds": 6.339064 }, { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 5.861859, "decode_tokens_per_second": 43.502, "finish_reason": "length", "measured_sse_events": 112, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 177, "reasoning_characters": 1120, "reasoning_sha256": "f8d00f6d3023d3fb97824661aa81d59973bc185ef4bdbfbc1be5c71ddba7a0e5", "response_tail_seconds": 0.000113, "stream_done_marker": true, "stream_sha256": "f8d00f6d3023d3fb97824661aa81d59973bc185ef4bdbfbc1be5c71ddba7a0e5", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 6.093883, "ttft_seconds": 0.232024, "wall_seconds": 6.093996 } ], "wall_seconds": 6.340284 }, { "aggregate_end_to_end_tokens_per_second": 86.28, "completion_tokens": 512, "round": 2, "streams": [ { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 5.423719, "decode_tokens_per_second": 47.016, "finish_reason": "length", "measured_sse_events": 114, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 182, "reasoning_characters": 1104, "reasoning_sha256": "ed06040e9cde60362a6ecc6f39f58ef13cec2f56569468f43aec45a6f5b91bbb", "response_tail_seconds": 8.5e-05, "stream_done_marker": true, "stream_sha256": "ed06040e9cde60362a6ecc6f39f58ef13cec2f56569468f43aec45a6f5b91bbb", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 5.932456, "ttft_seconds": 0.508736, "wall_seconds": 5.93254 }, { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 5.570902, "decode_tokens_per_second": 45.774, "finish_reason": "length", "measured_sse_events": 111, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 180, "reasoning_characters": 1128, "reasoning_sha256": "dce48a8b440ca04b78e27c848caa309eddfdce8e6f812e7c53115cc38955f7de", "response_tail_seconds": 0.0002, "stream_done_marker": true, "stream_sha256": "dce48a8b440ca04b78e27c848caa309eddfdce8e6f812e7c53115cc38955f7de", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 5.802191, "ttft_seconds": 0.23129, "wall_seconds": 5.802392 } ], "wall_seconds": 5.934197 }, { "aggregate_end_to_end_tokens_per_second": 77.184, "completion_tokens": 512, "round": 3, "streams": [ { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 6.14077, "decode_tokens_per_second": 41.526, "finish_reason": "length", "measured_sse_events": 125, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 181, "reasoning_characters": 1212, "reasoning_sha256": "684c507d75c1000320d0baaf6772392eff687009b3ebc8e9a42e40e3703193c4", "response_tail_seconds": 0.000122, "stream_done_marker": true, "stream_sha256": "684c507d75c1000320d0baaf6772392eff687009b3ebc8e9a42e40e3703193c4", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 6.631301, "ttft_seconds": 0.490531, "wall_seconds": 6.631423 }, { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 6.066437, "decode_tokens_per_second": 42.035, "finish_reason": "length", "measured_sse_events": 118, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 179, "reasoning_characters": 1102, "reasoning_sha256": "c650b7d5aea797261cd8f6b656b0cf5947f85532f6dc1134590f44332be49c6a", "response_tail_seconds": 0.000115, "stream_done_marker": true, "stream_sha256": "c650b7d5aea797261cd8f6b656b0cf5947f85532f6dc1134590f44332be49c6a", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 6.280701, "ttft_seconds": 0.214264, "wall_seconds": 6.280815 } ], "wall_seconds": 6.633477 } ] }, "4": { "aggregate_end_to_end_tokens_per_second": { "maximum": 127.382, "median": 125.493, "minimum": 123.813, "p90": 127.382 }, "per_stream_decode_tokens_per_second": { "maximum": 41.476, "median": 34.13, "minimum": 32.344, "p90": 36.798 }, "per_stream_ttft_seconds": { "maximum": 0.683, "median": 0.559, "minimum": 0.218, "p90": 0.683 }, "rounds": [ { "aggregate_end_to_end_tokens_per_second": 125.493, "completion_tokens": 1024, "round": 1, "streams": [ { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 7.796066, "decode_tokens_per_second": 32.709, "finish_reason": "length", "measured_sse_events": 119, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 181, "reasoning_characters": 1185, "reasoning_sha256": "1d739e799be19a3df03051d4e786ed0e487e7b969c345a04f6de3cb137f7edeb", "response_tail_seconds": 0.000128, "stream_done_marker": true, "stream_sha256": "1d739e799be19a3df03051d4e786ed0e487e7b969c345a04f6de3cb137f7edeb", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 8.013608, "ttft_seconds": 0.217542, "wall_seconds": 8.013737 }, { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 7.449981, "decode_tokens_per_second": 34.228, "finish_reason": "length", "measured_sse_events": 120, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 180, "reasoning_characters": 1133, "reasoning_sha256": "4c38dd3de88634219be98075d55ab592ced4a4d0388cdde0ee135fc24038dfc3", "response_tail_seconds": 0.000123, "stream_done_marker": true, "stream_sha256": "4c38dd3de88634219be98075d55ab592ced4a4d0388cdde0ee135fc24038dfc3", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 8.109614, "ttft_seconds": 0.659632, "wall_seconds": 8.109737 }, { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 7.495829, "decode_tokens_per_second": 34.019, "finish_reason": "length", "measured_sse_events": 121, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 179, "reasoning_characters": 1110, "reasoning_sha256": "1949a60edeae341fcce0c94e07c4fc8e6063b648bc5cec7b477ba263ceae0c72", "response_tail_seconds": 7.3e-05, "stream_done_marker": true, "stream_sha256": "1949a60edeae341fcce0c94e07c4fc8e6063b648bc5cec7b477ba263ceae0c72", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 8.156695, "ttft_seconds": 0.660866, "wall_seconds": 8.156767 }, { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 6.14815, "decode_tokens_per_second": 41.476, "finish_reason": "length", "measured_sse_events": 96, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 179, "reasoning_characters": 1067, "reasoning_sha256": "7a92d2e063d0c6c6ee3112eea15ede4ec90f2be3dc0cd10ddd7f6854191067bb", "response_tail_seconds": 0.000213, "stream_done_marker": true, "stream_sha256": "7a92d2e063d0c6c6ee3112eea15ede4ec90f2be3dc0cd10ddd7f6854191067bb", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 6.809258, "ttft_seconds": 0.661108, "wall_seconds": 6.80947 } ], "wall_seconds": 8.159793 }, { "aggregate_end_to_end_tokens_per_second": 123.813, "completion_tokens": 1024, "round": 2, "streams": [ { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 7.585784, "decode_tokens_per_second": 33.616, "finish_reason": "length", "measured_sse_events": 122, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 179, "reasoning_characters": 1167, "reasoning_sha256": "f4eef406d0abf0c6b17f50547452616d2cb967c4d8f23b80c576bbb1b311ee95", "response_tail_seconds": 0.000204, "stream_done_marker": true, "stream_sha256": "f4eef406d0abf0c6b17f50547452616d2cb967c4d8f23b80c576bbb1b311ee95", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 8.268114, "ttft_seconds": 0.68233, "wall_seconds": 8.268318 }, { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 7.26917, "decode_tokens_per_second": 35.08, "finish_reason": "length", "measured_sse_events": 116, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 180, "reasoning_characters": 1166, "reasoning_sha256": "2d8c593cd3ca259040ba9f1b4d1c4909f9f5eedc98a3509d8ea77a6d0c46ffed", "response_tail_seconds": 0.000113, "stream_done_marker": true, "stream_sha256": "2d8c593cd3ca259040ba9f1b4d1c4909f9f5eedc98a3509d8ea77a6d0c46ffed", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 7.952562, "ttft_seconds": 0.683392, "wall_seconds": 7.952674 }, { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 7.49277, "decode_tokens_per_second": 34.033, "finish_reason": "length", "measured_sse_events": 120, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 177, "reasoning_characters": 1157, "reasoning_sha256": "53a29ebccdceaa58797ca7b1b6ec9d9af767bd52cd64162bfbbea510173cc1b3", "response_tail_seconds": 0.000123, "stream_done_marker": true, "stream_sha256": "53a29ebccdceaa58797ca7b1b6ec9d9af767bd52cd64162bfbbea510173cc1b3", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 8.175767, "ttft_seconds": 0.682998, "wall_seconds": 8.17589 }, { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 7.883929, "decode_tokens_per_second": 32.344, "finish_reason": "length", "measured_sse_events": 120, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 181, "reasoning_characters": 1142, "reasoning_sha256": "eda3bb50df0ff5bd6b3d35c980323fa8e6eef4725b9a5476f94c0462ae04b9b2", "response_tail_seconds": 0.000189, "stream_done_marker": true, "stream_sha256": "eda3bb50df0ff5bd6b3d35c980323fa8e6eef4725b9a5476f94c0462ae04b9b2", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 8.116876, "ttft_seconds": 0.232947, "wall_seconds": 8.117065 } ], "wall_seconds": 8.270527 }, { "aggregate_end_to_end_tokens_per_second": 127.382, "completion_tokens": 1024, "round": 3, "streams": [ { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 7.577791, "decode_tokens_per_second": 33.651, "finish_reason": "length", "measured_sse_events": 121, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 179, "reasoning_characters": 1148, "reasoning_sha256": "dfb6076b25894e35bd432fa8c1f2b613b87d160d05010793516315a3ec135f92", "response_tail_seconds": 0.000152, "stream_done_marker": true, "stream_sha256": "dfb6076b25894e35bd432fa8c1f2b613b87d160d05010793516315a3ec135f92", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 8.036803, "ttft_seconds": 0.459012, "wall_seconds": 8.036955 }, { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 6.929723, "decode_tokens_per_second": 36.798, "finish_reason": "length", "measured_sse_events": 107, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 176, "reasoning_characters": 1136, "reasoning_sha256": "00591bc60f7e313463d664060b330021d60fa6a13bf47987aa2f5a52ec161a6e", "response_tail_seconds": 7.6e-05, "stream_done_marker": true, "stream_sha256": "00591bc60f7e313463d664060b330021d60fa6a13bf47987aa2f5a52ec161a6e", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 7.38907, "ttft_seconds": 0.459347, "wall_seconds": 7.389146 }, { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 7.172944, "decode_tokens_per_second": 35.55, "finish_reason": "length", "measured_sse_events": 111, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 183, "reasoning_characters": 1155, "reasoning_sha256": "e3ea2f4c4dfb67516c69f625ccb05c94508b7fcad09fbcb29ab6e984e29165a8", "response_tail_seconds": 0.000191, "stream_done_marker": true, "stream_sha256": "e3ea2f4c4dfb67516c69f625ccb05c94508b7fcad09fbcb29ab6e984e29165a8", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 7.631767, "ttft_seconds": 0.458822, "wall_seconds": 7.631958 }, { "cached_prompt_tokens": 0, "completion_tokens": 256, "decode_seconds": 7.173259, "decode_tokens_per_second": 35.549, "finish_reason": "length", "measured_sse_events": 111, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 179, "reasoning_characters": 1162, "reasoning_sha256": "4f5aeb2d82ab71411e4afe69ef211c5529e50a86e6123059c0bd8abdde9b91b3", "response_tail_seconds": 0.000139, "stream_done_marker": true, "stream_sha256": "4f5aeb2d82ab71411e4afe69ef211c5529e50a86e6123059c0bd8abdde9b91b3", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 7.632384, "ttft_seconds": 0.459125, "wall_seconds": 7.632523 } ], "wall_seconds": 8.038824 } ] } }, "decode": { "code": { "completion_gate": { "passed": 0, "total": 3 }, "decode_tokens_per_second": { "maximum": 131.715, "median": 87.4, "minimum": 75.253, "p90": 131.715 }, "pooled_decode_tokens_per_second": 92.815, "runs": [ { "cached_prompt_tokens": 0, "completion_tokens": 32768, "completion_validation": { "error": "no visible answer", "valid": false }, "decode_seconds": 248.772499, "decode_tokens_per_second": 131.715, "finish_reason": "length", "measured_sse_events": 4837, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 179, "reasoning_characters": 99431, "reasoning_sha256": "cde7f0dd6d81dbc5af315618513c715de4d35ae38b6b8cf818d3e14d685a689c", "response_tail_seconds": 0.002059, "run": 1, "stream_done_marker": true, "stream_sha256": "cde7f0dd6d81dbc5af315618513c715de4d35ae38b6b8cf818d3e14d685a689c", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 249.615131, "ttft_seconds": 0.842631, "wall_seconds": 249.61719 }, { "cached_prompt_tokens": 0, "completion_tokens": 32768, "completion_validation": { "error": "no visible answer", "valid": false }, "decode_seconds": 435.425284, "decode_tokens_per_second": 75.253, "finish_reason": "length", "measured_sse_events": 8283, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 182, "reasoning_characters": 119199, "reasoning_sha256": "9d2b78961bc92d687dfe725296d350bfa7548bcefa32f02984c9f431c7933494", "response_tail_seconds": 0.001812, "run": 2, "stream_done_marker": true, "stream_sha256": "9d2b78961bc92d687dfe725296d350bfa7548bcefa32f02984c9f431c7933494", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 435.736444, "ttft_seconds": 0.31116, "wall_seconds": 435.738256 }, { "cached_prompt_tokens": 0, "completion_tokens": 32768, "completion_validation": { "error": "no visible answer", "valid": false }, "decode_seconds": 374.909731, "decode_tokens_per_second": 87.4, "finish_reason": "length", "measured_sse_events": 7912, "output": "", "output_characters": 0, "output_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "prompt_tokens": 182, "reasoning_characters": 100972, "reasoning_sha256": "ac60a53b8cbca09e0ab707ea70c9e3c05488a86605b2300a330a6cb0c28c7213", "response_tail_seconds": 0.001608, "run": 3, "stream_done_marker": true, "stream_sha256": "ac60a53b8cbca09e0ab707ea70c9e3c05488a86605b2300a330a6cb0c28c7213", "time_to_first_visible_seconds": null, "time_to_last_output_seconds": 375.220737, "ttft_seconds": 0.311006, "wall_seconds": 375.222345 } ], "time_to_last_output_seconds": { "maximum": 435.736, "median": 375.221, "minimum": 249.615, "p90": 435.736 }, "ttft_seconds": { "maximum": 0.843, "median": 0.311, "minimum": 0.311, "p90": 0.843 }, "wall_seconds": { "maximum": 435.738, "median": 375.222, "minimum": 249.617, "p90": 435.738 } }, "prose": { "completion_gate": { "passed": 3, "total": 3 }, "decode_tokens_per_second": { "maximum": 113.222, "median": 93.98, "minimum": 82.596, "p90": 113.222 }, "pooled_decode_tokens_per_second": 101.459, "runs": [ { "cached_prompt_tokens": 0, "completion_tokens": 25223, "completion_validation": { "valid": true }, "decode_seconds": 222.766053, "decode_tokens_per_second": 113.222, "finish_reason": "stop", "measured_sse_events": 4104, "output": "MEMORANDUM\n\nTo: Engineering Team \nFrom: Infrastructure Lead \nSubject: Local AI hardware: buy more or improve utilization\n\nWe are at a crossroads: the team feels pressure to add local AI hardware, while the counterargument is that we should first get more value from the machines we already own. The decision should be framed as marginal benefit: whether the next dollar spent on hardware produces more useful capacity, lower latency, or better reliability than the next dollar spent on engineering effort.\n\nOur evidence is mixed. Some jobs queue during peak hours, suggesting capacity is tight. Other machines sit idle overnight, suggesting capacity is not uniformly scarce. Until we separate these patterns, we cannot know whether the problem is a shortage of silicon or of scheduling, batching, model loading, and observability. A purchase may be ineffective if the real bottleneck is memory bandwidth, cold starts, or poorly instrumented queues.\n\nCost is the first place to apply that discipline. A local GPU purchase is not just the invoice. It is power, cooling, rack space, and the engineering time required to integrate new nodes into our deployment processes. If utilization is low, we are paying to keep expensive hardware warm. If utilization is high, the purchase may be justified, but only if the added capacity is actually consumed. Buying hardware now may postpone software work that could make the fleet cheaper to run for years.\n\nLocal hardware can reduce network latency and avoid cloud queueing, a real advantage for interactive inference. However, if our existing machines are saturated, adding more hardware may not improve tail latency unless the scheduler can use it well. A poorly balanced cluster can produce worse p95 than a smaller, well-tuned one. Better batching, model caching, and workload placement may reduce queue time without buying anything. The right metric is the latency distribution experienced by users and the cost per successful request.\n\nReliability cuts both ways. More hardware can provide redundancy and spare capacity, valuable when a node fails or a model must be served during a traffic spike. But it also introduces more failure domains and operational surface area. A small team cannot always afford to operate a larger fleet. Sometimes the most reliable system is the one whose behavior is understood and whose capacity can be shifted without emergency changes. If we buy hardware, we should be confident we can operate it without adding fragility.\n\nOpportunity cost is the strongest argument for improving utilization first. Every week spent integrating hardware is a week not spent reducing model latency, improving quantization, or building better routing. If we can raise effective throughput by twenty or thirty percent through software and scheduling, that improvement compounds across every future workload. Hardware is a one-time addition; it may be necessary, but it is not always the most productive first move.\n\nReversibility should also weigh heavily. A hardware purchase is difficult to undo, and its value declines as models and vendor roadmaps change. A utilization program is more reversible: if a batching change hurts quality, we can roll it back; if a scheduling policy is wrong, we can adjust it. Software improvements are easier to tune and less likely to lock us into a particular architecture. That favors starting with the work we can control.\n\nMy recommendation is to defer a hardware purchase for the next six to eight weeks and run a focused utilization effort. We should instrument queue depth, GPU utilization, memory pressure, batch efficiency, and p95 latency by workload. We should then make targeted changes such as better batching, model caching, and workload placement. If, after that effort, we still see sustained utilization above eighty-five percent and p95 latency outside our targets, we should buy hardware, but only for the specific bottleneck identified. If utilization remains below sixty percent, we should not buy. If the picture is mixed, buy only enough to cover the peak workload, not the average.\n\nThis recommendation is not a claim that hardware is unnecessary. It is a claim that we should not buy capacity before we understand the shape of the constraint. Demand and cloud alternatives may change. We should revisit the decision after the utilization work. For now, the most prudent engineering move is to improve what we own, measure honestly, and keep the option to buy open.", "output_characters": 4485, "output_sha256": "70e5304dfdc989e74e0465af3e8c2dbac9c71752873eb195c7643c16172f08a0", "prompt_tokens": 175, "reasoning_characters": 86806, "reasoning_sha256": "c926b99dd05412302f9073bed31d050257433a1e853c2e8f681d5860ee515905", "response_tail_seconds": 0.149664, "run": 1, "stream_done_marker": true, "stream_sha256": "331ed93396764f5566fc0e5e25af7c9cc8fde8e12eaae94b8a6dbb1b3f532dd4", "time_to_first_visible_seconds": 217.944319, "time_to_last_output_seconds": 223.280956, "ttft_seconds": 0.514903, "wall_seconds": 223.43062 }, { "cached_prompt_tokens": 0, "completion_tokens": 5157, "completion_validation": { "valid": true }, "decode_seconds": 62.424164, "decode_tokens_per_second": 82.596, "finish_reason": "stop", "measured_sse_events": 1285, "output": "To: Engineering Team \nFrom: Infrastructure Lead \nSubject: Whether to buy more local AI hardware or improve utilisation of what we own\n\nWe need to decide whether the next major infrastructure move is to purchase additional local AI hardware or to invest first in using the machines we already have more effectively. This is not a simple capacity question. It is a question about how we spend scarce capital, how we protect user experience, and how much future flexibility we want to preserve. The honest answer is that we do not yet have enough evidence to justify a large hardware purchase, and that the most prudent path is to improve utilisation before adding more boxes.\n\nOur current fleet is probably not as idle as it feels, but it is also not as saturated as our most urgent incidents suggest. During peak inference traffic, some services queue badly and latency rises, while during nights and weekends many GPUs sit underused. That pattern points less to a shortage of silicon than to a shortage of scheduling, batching, caching, and workload isolation. If we buy more hardware now, we may smooth the next incident, but we will also lock in depreciation, power, cooling, maintenance windows, and operational complexity. Those costs are real and they do not disappear when the next model update makes the hardware less efficient.\n\nLatency is the strongest argument for local capacity, but it is not automatically an argument for more capacity. A request can be slow because the model is large, because the serving stack is inefficient, because the GPU is shared with noisy batch jobs, or because the request is waiting behind a poorly prioritised queue. Adding another node can help if the bottleneck is genuinely physical, but it can also make the system less predictable by increasing the number of places where state, model versions, and failure modes must be coordinated. Reliability cuts both ways. More hardware can provide redundancy and reduce blast radius, yet it also increases the number of components that can fail, the number of drivers and firmware versions that must be patched, and the number of capacity hotspots that can silently develop.\n\nThe opportunity cost is the part that deserves the most weight. Every dollar and every engineering week spent selecting, purchasing, racking, and integrating new hardware is a dollar and a week not spent making our existing platform better. We have a backlog of product work, reliability work, and model-serving improvements that could reduce cost and latency without capital expenditure. If we buy hardware, we are not merely buying compute; we are buying a future obligation to keep it useful. If we improve utilisation, we are buying optionality. We can always buy later if demand proves durable, but we cannot easily unbuy hardware once the market, our models, or our product needs change.\n\nThat reversibility matters. A hardware purchase is hard to reverse. It commits us to a particular vendor, architecture, and depreciation schedule, and it can bias us toward using the hardware simply because we own it. Utilisation work is easier to reverse. We can tune schedulers, adjust autoscaling policies, change model formats, and experiment with caching or quantization, then roll back if a change does not help. This is not a reason to avoid all purchases forever, but it is a reason to prefer reversible actions when uncertainty is high.\n\nMy recommendation is to spend the next quarter improving utilisation of the existing fleet and to defer any major hardware purchase until we have clearer evidence. We should build better observability around GPU utilisation, queue depth, cost per request, and end-to-end latency. We should separate interactive inference from batch training and experimentation, improve batching and model serving, and set explicit thresholds for when additional capacity is justified. If, after those changes, we consistently see sustained utilisation above a high threshold, miss latency targets, and cannot close the gap through software, then we should buy a measured amount of hardware.\n\nI acknowledge uncertainty. Demand may grow faster than expected, model requirements may change, and the economics of cloud or local hardware may shift. A small, targeted purchase could be reasonable if it removes a known bottleneck. But the burden of proof should be on the hardware case, not the utilisation case. For now, the better engineering bet is to make what we own work harder.", "output_characters": 4470, "output_sha256": "3f019d3111173beeb4885b8e03b43848ea657e7c87fc3851f137b3fca7f5afec", "prompt_tokens": 172, "reasoning_characters": 13288, "reasoning_sha256": "33c824c23cd2ebec0777d21dfa5e35cd179abe4405ef95d1fbabe4e1bc013226", "response_tail_seconds": 0.000617, "run": 2, "stream_done_marker": true, "stream_sha256": "424babbf7ad44a87ada74e93479e51e3e637239fd4c6adefc3a73850de9ce67b", "time_to_first_visible_seconds": 59.483771, "time_to_last_output_seconds": 62.661797, "ttft_seconds": 0.237633, "wall_seconds": 62.662414 }, { "cached_prompt_tokens": 0, "completion_tokens": 18134, "completion_validation": { "valid": true }, "decode_seconds": 192.944613, "decode_tokens_per_second": 93.98, "finish_reason": "stop", "measured_sse_events": 3779, "output": "To: Engineering Team\nFrom: Platform Engineering\nDate: 2026-06-15\nSubject: Local AI hardware versus utilization\nReference: be9ccb40e186ce6cb9d5cbb977b0354e\n\nWe need decide whether to buy more local AI hardware or improve utilization of what we own. The question is not whether GPUs are busy, but whether an extra machine buys something we cannot obtain more cheaply through scheduling, efficiency, and discipline. Our fleet is not idle, but unevenly used: some jobs queue while others hold capacity, and services run below target because autoscaling is conservative and batch workloads are poorly packed. Before adding hardware, we should prove the bottleneck is capacity, not coordination.\n\nCost is the most visible factor. A new node is not just the purchase price; it is power, cooling, rack space, maintenance, licensing, and engineering time. If better packing, priority queues, and quantization raise utilization by twenty or thirty percent, we may defer a purchase for months. That deferral is valuable because hardware prices fall and software improves. The opportunity cost of buying now is not only cash, but the risk of locking ourselves into an obsolete configuration before it pays for itself.\n\nLatency matters too. If interactive workloads are slow, we need to separate capacity latency from scheduling latency. A request waiting behind a long training job is scheduling, not hardware. A request that is slow because the model is large or the stack is inefficient is software. Adding GPUs may reduce queue time, but it will not make a poorly optimized inference path faster. Separating interactive and batch workloads, enforcing fair-share policies, and using smaller models where quality permits may improve perceived latency without buying anything.\n\nReliability is another reason to be cautious. More hardware means more failure domains, drivers, firmware, monitoring, and complexity. A small team cannot afford to firefight a larger cluster. Improving utilization can also improve reliability by making the system easier to reason about, with fewer nodes, clearer policies, and better observability. But if utilization is already too high, reliability suffers because there is no headroom for failures, retries, or spikes. The target is not maximum utilization, but sustainable utilization with slack to absorb normal variance.\n\nOpportunity cost cuts both ways. Every engineer-hour spent tuning schedulers, profiling models, or building autoscaling is an hour not spent shipping. If long queues block the team and the business is losing revenue or velocity, waiting may be more expensive than buying. Then a modest purchase may be justified to buy time. But we should be honest about whether the blockage is real or a symptom of poor visibility and weak habits. If we cannot say where time goes, we are not ready to buy.\n\nReversibility should guide the decision. A hardware purchase is hard to reverse: it commits us to depreciation, maintenance, and an architecture. A utilization improvement is more reversible; if it fails, we can still buy hardware later with better data. This asymmetry favors the cheaper, more reversible option. We should not treat this decision as permanent. We can set a clear review point, such as eight weeks, and define when we would change course.\n\nMy recommendation is to improve utilization first, but not indefinitely. We should implement a focused effort over the next quarter, including better job classification, priority queues for interactive workloads, batch packing, autoscaling based on queue depth and latency, and profiling of the most expensive models. We should also add dashboards showing utilization, queue time, failure headroom, and cost per successful request. If, after this work, interactive latency remains above target, queue time is structurally high, and we have no reliable way to reduce model cost, then we should buy a limited amount of hardware, preferably a small incremental cluster that can be tested before full commitment.\n\nThere is uncertainty here. We may discover that our models are already efficient and the fleet is genuinely saturated. We may also discover that the real problem is not hardware, but poor data pipelines, inefficient code, or unclear priorities. The memo should not pretend we know the answer. What we do know is that buying more hardware is the more expensive and less reversible choice, and that our current evidence does not yet justify it. We should buy information first, then hardware only if the data demands it.", "output_characters": 4524, "output_sha256": "f2f9b43ae32a81885d296d0f7d57a24674baecf377f3b509838c1b1d58de1f9f", "prompt_tokens": 168, "reasoning_characters": 63543, "reasoning_sha256": "ad75e8ab3f6dcdbd7a80c1d9cfd1a92d775d553df08e55196b6f99cd134b03d4", "response_tail_seconds": 0.001216, "run": 3, "stream_done_marker": true, "stream_sha256": "a4ea434677e098790bda74a91fbda3ac2f304ac22a5e7a81b4c22e314b2f1521", "time_to_first_visible_seconds": 189.736908, "time_to_last_output_seconds": 193.181169, "ttft_seconds": 0.236556, "wall_seconds": 193.182385 } ], "time_to_last_output_seconds": { "maximum": 223.281, "median": 193.181, "minimum": 62.662, "p90": 223.281 }, "ttft_seconds": { "maximum": 0.515, "median": 0.238, "minimum": 0.237, "p90": 0.515 }, "wall_seconds": { "maximum": 223.431, "median": 193.182, "minimum": 62.662, "p90": 223.431 } }, "structured": { "completion_gate": { "passed": 3, "total": 3 }, "decode_tokens_per_second": { "maximum": 141.016, "median": 140.133, "minimum": 118.744, "p90": 141.016 }, "pooled_decode_tokens_per_second": 130.178, "runs": [ { "cached_prompt_tokens": 0, "completion_tokens": 797, "completion_validation": { "valid": true }, "decode_seconds": 5.644753, "decode_tokens_per_second": 141.016, "finish_reason": "stop", "measured_sse_events": 80, "output": "[\n {\"index\":1,\"square\":1},\n {\"index\":2,\"square\":4},\n {\"index\":3,\"square\":9},\n {\"index\":4,\"square\":16},\n {\"index\":5,\"square\":25},\n {\"index\":6,\"square\":36},\n {\"index\":7,\"square\":49},\n {\"index\":8,\"square\":64},\n {\"index\":9,\"square\":81},\n {\"index\":10,\"square\":100},\n {\"index\":11,\"square\":121},\n {\"index\":12,\"square\":144},\n {\"index\":13,\"square\":169},\n {\"index\":14,\"square\":196},\n {\"index\":15,\"square\":225},\n {\"index\":16,\"square\":256},\n {\"index\":17,\"square\":289},\n {\"index\":18,\"square\":324},\n {\"index\":19,\"square\":361},\n {\"index\":20,\"square\":400},\n {\"index\":21,\"square\":441},\n {\"index\":22,\"square\":484},\n {\"index\":23,\"square\":529},\n {\"index\":24,\"square\":576},\n {\"index\":25,\"square\":625},\n {\"index\":26,\"square\":676},\n {\"index\":27,\"square\":729},\n {\"index\":28,\"square\":784},\n {\"index\":29,\"square\":841},\n {\"index\":30,\"square\":900},\n {\"index\":31,\"square\":961},\n {\"index\":32,\"square\":1024},\n {\"index\":33,\"square\":1089},\n {\"index\":34,\"square\":1156},\n {\"index\":35,\"square\":1225},\n {\"index\":36,\"square\":1296},\n {\"index\":37,\"square\":1369},\n {\"index\":38,\"square\":1444},\n {\"index\":39,\"square\":1521},\n {\"index\":40,\"square\":1600},\n {\"index\":41,\"square\":1681},\n {\"index\":42,\"square\":1764},\n {\"index\":43,\"square\":1849},\n {\"index\":44,\"square\":1936},\n {\"index\":45,\"square\":2025},\n {\"index\":46,\"square\":2116},\n {\"index\":47,\"square\":2209},\n {\"index\":48,\"square\":2304},\n {\"index\":49,\"square\":2401},\n {\"index\":50,\"square\":2500}\n]", "output_characters": 1450, "output_sha256": "ec1727e6c0adbdd1452bff29c15d3802874e5d068da31595184706d022187865", "prompt_tokens": 160, "reasoning_characters": 414, "reasoning_sha256": "4b709c80909c828b6b681b4865fd35ba9dc8577688c08511ca62a7eb16c69ac4", "response_tail_seconds": 0.000116, "run": 1, "stream_done_marker": true, "stream_sha256": "24883e8a5508ecb563fe4b2a25c13f5ccd19fc86df5a2eb4c535762328bb4c26", "time_to_first_visible_seconds": 1.459583, "time_to_last_output_seconds": 5.939511, "ttft_seconds": 0.294757, "wall_seconds": 5.939626 }, { "cached_prompt_tokens": 0, "completion_tokens": 1186, "completion_validation": { "valid": true }, "decode_seconds": 9.979465, "decode_tokens_per_second": 118.744, "finish_reason": "stop", "measured_sse_events": 168, "output": "[\n {\"index\":1,\"square\":1},\n {\"index\":2,\"square\":4},\n {\"index\":3,\"square\":9},\n {\"index\":4,\"square\":16},\n {\"index\":5,\"square\":25},\n {\"index\":6,\"square\":36},\n {\"index\":7,\"square\":49},\n {\"index\":8,\"square\":64},\n {\"index\":9,\"square\":81},\n {\"index\":10,\"square\":100},\n {\"index\":11,\"square\":121},\n {\"index\":12,\"square\":144},\n {\"index\":13,\"square\":169},\n {\"index\":14,\"square\":196},\n {\"index\":15,\"square\":225},\n {\"index\":16,\"square\":256},\n {\"index\":17,\"square\":289},\n {\"index\":18,\"square\":324},\n {\"index\":19,\"square\":361},\n {\"index\":20,\"square\":400},\n {\"index\":21,\"square\":441},\n {\"index\":22,\"square\":484},\n {\"index\":23,\"square\":529},\n {\"index\":24,\"square\":576},\n {\"index\":25,\"square\":625},\n {\"index\":26,\"square\":676},\n {\"index\":27,\"square\":729},\n {\"index\":28,\"square\":784},\n {\"index\":29,\"square\":841},\n {\"index\":30,\"square\":900},\n {\"index\":31,\"square\":961},\n {\"index\":32,\"square\":1024},\n {\"index\":33,\"square\":1089},\n {\"index\":34,\"square\":1156},\n {\"index\":35,\"square\":1225},\n {\"index\":36,\"square\":1296},\n {\"index\":37,\"square\":1369},\n {\"index\":38,\"square\":1444},\n {\"index\":39,\"square\":1521},\n {\"index\":40,\"square\":1600},\n {\"index\":41,\"square\":1681},\n {\"index\":42,\"square\":1764},\n {\"index\":43,\"square\":1849},\n {\"index\":44,\"square\":1936},\n {\"index\":45,\"square\":2025},\n {\"index\":46,\"square\":2116},\n {\"index\":47,\"square\":2209},\n {\"index\":48,\"square\":2304},\n {\"index\":49,\"square\":2401},\n {\"index\":50,\"square\":2500}\n]", "output_characters": 1450, "output_sha256": "ec1727e6c0adbdd1452bff29c15d3802874e5d068da31595184706d022187865", "prompt_tokens": 159, "reasoning_characters": 870, "reasoning_sha256": "b179dfa7bf1dde35a0a87c57c2749a8e339d82493cbdee1773cccb25a5e7f5cf", "response_tail_seconds": 7.4e-05, "run": 2, "stream_done_marker": true, "stream_sha256": "6a6f7d5f03c0f0692e1d6c78b748c6ca393c392d94e8c54ac2412cc6a26c2f4f", "time_to_first_visible_seconds": 6.030693, "time_to_last_output_seconds": 10.209197, "ttft_seconds": 0.229732, "wall_seconds": 10.209272 }, { "cached_prompt_tokens": 0, "completion_tokens": 746, "completion_validation": { "valid": true }, "decode_seconds": 5.316393, "decode_tokens_per_second": 140.133, "finish_reason": "stop", "measured_sse_events": 81, "output": "[\n {\"index\":1,\"square\":1},\n {\"index\":2,\"square\":4},\n {\"index\":3,\"square\":9},\n {\"index\":4,\"square\":16},\n {\"index\":5,\"square\":25},\n {\"index\":6,\"square\":36},\n {\"index\":7,\"square\":49},\n {\"index\":8,\"square\":64},\n {\"index\":9,\"square\":81},\n {\"index\":10,\"square\":100},\n {\"index\":11,\"square\":121},\n {\"index\":12,\"square\":144},\n {\"index\":13,\"square\":169},\n {\"index\":14,\"square\":196},\n {\"index\":15,\"square\":225},\n {\"index\":16,\"square\":256},\n {\"index\":17,\"square\":289},\n {\"index\":18,\"square\":324},\n {\"index\":19,\"square\":361},\n {\"index\":20,\"square\":400},\n {\"index\":21,\"square\":441},\n {\"index\":22,\"square\":484},\n {\"index\":23,\"square\":529},\n {\"index\":24,\"square\":576},\n {\"index\":25,\"square\":625},\n {\"index\":26,\"square\":676},\n {\"index\":27,\"square\":729},\n {\"index\":28,\"square\":784},\n {\"index\":29,\"square\":841},\n {\"index\":30,\"square\":900},\n {\"index\":31,\"square\":961},\n {\"index\":32,\"square\":1024},\n {\"index\":33,\"square\":1089},\n {\"index\":34,\"square\":1156},\n {\"index\":35,\"square\":1225},\n {\"index\":36,\"square\":1296},\n {\"index\":37,\"square\":1369},\n {\"index\":38,\"square\":1444},\n {\"index\":39,\"square\":1521},\n {\"index\":40,\"square\":1600},\n {\"index\":41,\"square\":1681},\n {\"index\":42,\"square\":1764},\n {\"index\":43,\"square\":1849},\n {\"index\":44,\"square\":1936},\n {\"index\":45,\"square\":2025},\n {\"index\":46,\"square\":2116},\n {\"index\":47,\"square\":2209},\n {\"index\":48,\"square\":2304},\n {\"index\":49,\"square\":2401},\n {\"index\":50,\"square\":2500}\n]", "output_characters": 1450, "output_sha256": "ec1727e6c0adbdd1452bff29c15d3802874e5d068da31595184706d022187865", "prompt_tokens": 162, "reasoning_characters": 194, "reasoning_sha256": "853aaf3db11a95b0c11e256a1d3a7565c8ee2ae8b918e1f75f57e51c7aabd8cc", "response_tail_seconds": 0.000234, "run": 3, "stream_done_marker": true, "stream_sha256": "67c66d1e3be573f9d719147ed49a6561a15c64a060a40fe3e221c347e02fab68", "time_to_first_visible_seconds": 1.163124, "time_to_last_output_seconds": 5.542373, "ttft_seconds": 0.22598, "wall_seconds": 5.542607 } ], "time_to_last_output_seconds": { "maximum": 10.209, "median": 5.94, "minimum": 5.542, "p90": 10.209 }, "ttft_seconds": { "maximum": 0.295, "median": 0.23, "minimum": 0.226, "p90": 0.295 }, "wall_seconds": { "maximum": 10.209, "median": 5.94, "minimum": 5.543, "p90": 10.209 } } }, "prefill": { "32768": { "cold": { "effective_prefill_tokens_per_second": { "maximum": 2520.5, "median": 2518.404, "minimum": 2516.853, "p90": 2520.5 }, "runs": [ { "cache_salt_sha256": "43815bd6a8dc5258ba1dbc5185324551c9a0e2ecb19f1bb46302becddda5fc55", "cached_prompt_tokens": 0, "completion_tokens": 8, "decode_seconds": 0.085745, "decode_tokens_per_second": 81.637, "effective_prefill_tokens_per_second": 2518.404, "finish_reason": "length", "measured_sse_events": 2, "output": " behaviour, evidence, constraints, and the", "output_characters": 42, "output_sha256": "128002ef37a31be1bbdb67225d75ef3c0c801e9fa83861d8e04741bdff280850", "prompt_tokens": 32768, "reasoning_characters": 0, "reasoning_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "requested_prompt_tokens": 32768, "response_tail_seconds": 0.000249, "run": 1, "stream_done_marker": true, "stream_sha256": "128002ef37a31be1bbdb67225d75ef3c0c801e9fa83861d8e04741bdff280850", "time_to_first_visible_seconds": 13.011415, "time_to_last_output_seconds": 13.09716, "ttft_seconds": 13.011415, "wall_seconds": 13.097409 }, { "cache_salt_sha256": "2de85f70aea6c1888b47a435651c872ad12cc2d788a5c6a539b6730dc9610dc0", "cached_prompt_tokens": 0, "completion_tokens": 8, "decode_seconds": 0.085562, "decode_tokens_per_second": 81.812, "effective_prefill_tokens_per_second": 2520.5, "finish_reason": "length", "measured_sse_events": 2, "output": " behaviour, evidence, constraints, and the", "output_characters": 42, "output_sha256": "128002ef37a31be1bbdb67225d75ef3c0c801e9fa83861d8e04741bdff280850", "prompt_tokens": 32768, "reasoning_characters": 0, "reasoning_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "requested_prompt_tokens": 32768, "response_tail_seconds": 0.000175, "run": 2, "stream_done_marker": true, "stream_sha256": "128002ef37a31be1bbdb67225d75ef3c0c801e9fa83861d8e04741bdff280850", "time_to_first_visible_seconds": 13.000596, "time_to_last_output_seconds": 13.086158, "ttft_seconds": 13.000596, "wall_seconds": 13.086333 }, { "cache_salt_sha256": "c9b53fb6ddfba531e5f4a990ef42e29eaf66422528136b829bdcd82fecb4f1ea", "cached_prompt_tokens": 0, "completion_tokens": 8, "decode_seconds": 0.085149, "decode_tokens_per_second": 82.209, "effective_prefill_tokens_per_second": 2516.853, "finish_reason": "length", "measured_sse_events": 2, "output": " behaviour, evidence, constraints, and the", "output_characters": 42, "output_sha256": "128002ef37a31be1bbdb67225d75ef3c0c801e9fa83861d8e04741bdff280850", "prompt_tokens": 32768, "reasoning_characters": 0, "reasoning_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "requested_prompt_tokens": 32768, "response_tail_seconds": 0.000162, "run": 3, "stream_done_marker": true, "stream_sha256": "128002ef37a31be1bbdb67225d75ef3c0c801e9fa83861d8e04741bdff280850", "time_to_first_visible_seconds": 13.019435, "time_to_last_output_seconds": 13.104584, "ttft_seconds": 13.019435, "wall_seconds": 13.104747 } ], "ttft_seconds": { "maximum": 13.019, "median": 13.011, "minimum": 13.001, "p90": 13.019 } }, "warm_replay": { "effective_prefill_tokens_per_second": { "maximum": 2529.525, "median": 2524.007, "minimum": 2520.473, "p90": 2529.525 }, "runs": [ { "cache_salt_sha256": "43815bd6a8dc5258ba1dbc5185324551c9a0e2ecb19f1bb46302becddda5fc55", "cached_prompt_tokens": 0, "completion_tokens": 8, "decode_seconds": 0.087216, "decode_tokens_per_second": 80.26, "effective_prefill_tokens_per_second": 2529.525, "finish_reason": "length", "measured_sse_events": 2, "output": " behaviour, evidence, constraints, and the", "output_characters": 42, "output_sha256": "128002ef37a31be1bbdb67225d75ef3c0c801e9fa83861d8e04741bdff280850", "prompt_tokens": 32768, "reasoning_characters": 0, "reasoning_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "requested_prompt_tokens": 32768, "response_tail_seconds": 0.000167, "run": 1, "stream_done_marker": true, "stream_sha256": "128002ef37a31be1bbdb67225d75ef3c0c801e9fa83861d8e04741bdff280850", "time_to_first_visible_seconds": 12.954213, "time_to_last_output_seconds": 13.041429, "ttft_seconds": 12.954213, "wall_seconds": 13.041596 }, { "cache_salt_sha256": "2de85f70aea6c1888b47a435651c872ad12cc2d788a5c6a539b6730dc9610dc0", "cached_prompt_tokens": 0, "completion_tokens": 8, "decode_seconds": 0.084098, "decode_tokens_per_second": 83.236, "effective_prefill_tokens_per_second": 2520.473, "finish_reason": "length", "measured_sse_events": 2, "output": " behaviour, evidence, constraints, and the", "output_characters": 42, "output_sha256": "128002ef37a31be1bbdb67225d75ef3c0c801e9fa83861d8e04741bdff280850", "prompt_tokens": 32768, "reasoning_characters": 0, "reasoning_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "requested_prompt_tokens": 32768, "response_tail_seconds": 0.000214, "run": 2, "stream_done_marker": true, "stream_sha256": "128002ef37a31be1bbdb67225d75ef3c0c801e9fa83861d8e04741bdff280850", "time_to_first_visible_seconds": 13.000736, "time_to_last_output_seconds": 13.084834, "ttft_seconds": 13.000736, "wall_seconds": 13.085048 }, { "cache_salt_sha256": "c9b53fb6ddfba531e5f4a990ef42e29eaf66422528136b829bdcd82fecb4f1ea", "cached_prompt_tokens": 0, "completion_tokens": 8, "decode_seconds": 0.084719, "decode_tokens_per_second": 82.626, "effective_prefill_tokens_per_second": 2524.007, "finish_reason": "length", "measured_sse_events": 2, "output": " behaviour, evidence, constraints, and the", "output_characters": 42, "output_sha256": "128002ef37a31be1bbdb67225d75ef3c0c801e9fa83861d8e04741bdff280850", "prompt_tokens": 32768, "reasoning_characters": 0, "reasoning_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "requested_prompt_tokens": 32768, "response_tail_seconds": 0.000162, "run": 3, "stream_done_marker": true, "stream_sha256": "128002ef37a31be1bbdb67225d75ef3c0c801e9fa83861d8e04741bdff280850", "time_to_first_visible_seconds": 12.98253, "time_to_last_output_seconds": 13.067249, "ttft_seconds": 12.98253, "wall_seconds": 13.067411 } ], "ttft_seconds": { "maximum": 13.001, "median": 12.983, "minimum": 12.954, "p90": 13.001 } } }, "65536": { "cold": { "effective_prefill_tokens_per_second": { "maximum": 2436.194, "median": 2434.418, "minimum": 2434.408, "p90": 2436.194 }, "runs": [ { "cache_salt_sha256": "0494a014140990ebb884e8aa4cd4067979ffe1bf45a4fd620ee45c3be1f7f257", "cached_prompt_tokens": 0, "completion_tokens": 8, "decode_seconds": 0.103271, "decode_tokens_per_second": 67.783, "effective_prefill_tokens_per_second": 2434.418, "finish_reason": "length", "measured_sse_events": 2, "output": " Each observation records behaviour, evidence, constraints", "output_characters": 58, "output_sha256": "3f91e56380da40aa484278ca6ad46278f1415f32fb168c9a22b4e5fb220e39c0", "prompt_tokens": 65536, "reasoning_characters": 0, "reasoning_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "requested_prompt_tokens": 65536, "response_tail_seconds": 0.000164, "run": 1, "stream_done_marker": true, "stream_sha256": "3f91e56380da40aa484278ca6ad46278f1415f32fb168c9a22b4e5fb220e39c0", "time_to_first_visible_seconds": 26.920598, "time_to_last_output_seconds": 27.023869, "ttft_seconds": 26.920598, "wall_seconds": 27.024034 }, { "cache_salt_sha256": "457716fada2fa637a812d9867cbf5d9d7c249f4f4096cf0d75440196fbc70d50", "cached_prompt_tokens": 0, "completion_tokens": 8, "decode_seconds": 0.10324, "decode_tokens_per_second": 67.803, "effective_prefill_tokens_per_second": 2434.408, "finish_reason": "length", "measured_sse_events": 2, "output": " observation records behaviour, evidence, constraints,", "output_characters": 54, "output_sha256": "64de6b20e549222d7456ec41b7e90d976a3186746092d94e04bb350c64dbcfd2", "prompt_tokens": 65536, "reasoning_characters": 0, "reasoning_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "requested_prompt_tokens": 65536, "response_tail_seconds": 0.000171, "run": 2, "stream_done_marker": true, "stream_sha256": "64de6b20e549222d7456ec41b7e90d976a3186746092d94e04bb350c64dbcfd2", "time_to_first_visible_seconds": 26.92071, "time_to_last_output_seconds": 27.02395, "ttft_seconds": 26.92071, "wall_seconds": 27.024121 }, { "cache_salt_sha256": "e4309e295fbeb64d79c97adc4cf08618024d772d3f79d6ed878b89064d172961", "cached_prompt_tokens": 0, "completion_tokens": 8, "decode_seconds": 0.103808, "decode_tokens_per_second": 67.432, "effective_prefill_tokens_per_second": 2436.194, "finish_reason": "length", "measured_sse_events": 2, "output": " Each observation records behaviour, evidence, constraints", "output_characters": 58, "output_sha256": "3f91e56380da40aa484278ca6ad46278f1415f32fb168c9a22b4e5fb220e39c0", "prompt_tokens": 65536, "reasoning_characters": 0, "reasoning_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "requested_prompt_tokens": 65536, "response_tail_seconds": 0.000168, "run": 3, "stream_done_marker": true, "stream_sha256": "3f91e56380da40aa484278ca6ad46278f1415f32fb168c9a22b4e5fb220e39c0", "time_to_first_visible_seconds": 26.900982, "time_to_last_output_seconds": 27.00479, "ttft_seconds": 26.900982, "wall_seconds": 27.004959 } ], "ttft_seconds": { "maximum": 26.921, "median": 26.921, "minimum": 26.901, "p90": 26.921 } }, "warm_replay": { "effective_prefill_tokens_per_second": { "maximum": 2441.801, "median": 2439.649, "minimum": 2437.65, "p90": 2441.801 }, "runs": [ { "cache_salt_sha256": "0494a014140990ebb884e8aa4cd4067979ffe1bf45a4fd620ee45c3be1f7f257", "cached_prompt_tokens": 0, "completion_tokens": 8, "decode_seconds": 0.104141, "decode_tokens_per_second": 67.217, "effective_prefill_tokens_per_second": 2437.65, "finish_reason": "length", "measured_sse_events": 2, "output": " Each observation records behaviour, evidence, constraints", "output_characters": 58, "output_sha256": "3f91e56380da40aa484278ca6ad46278f1415f32fb168c9a22b4e5fb220e39c0", "prompt_tokens": 65536, "reasoning_characters": 0, "reasoning_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "requested_prompt_tokens": 65536, "response_tail_seconds": 0.000166, "run": 1, "stream_done_marker": true, "stream_sha256": "3f91e56380da40aa484278ca6ad46278f1415f32fb168c9a22b4e5fb220e39c0", "time_to_first_visible_seconds": 26.884906, "time_to_last_output_seconds": 26.989047, "ttft_seconds": 26.884906, "wall_seconds": 26.989213 }, { "cache_salt_sha256": "457716fada2fa637a812d9867cbf5d9d7c249f4f4096cf0d75440196fbc70d50", "cached_prompt_tokens": 0, "completion_tokens": 8, "decode_seconds": 0.103238, "decode_tokens_per_second": 67.804, "effective_prefill_tokens_per_second": 2441.801, "finish_reason": "length", "measured_sse_events": 2, "output": " observation records behaviour, evidence, constraints,", "output_characters": 54, "output_sha256": "64de6b20e549222d7456ec41b7e90d976a3186746092d94e04bb350c64dbcfd2", "prompt_tokens": 65536, "reasoning_characters": 0, "reasoning_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "requested_prompt_tokens": 65536, "response_tail_seconds": 0.000168, "run": 2, "stream_done_marker": true, "stream_sha256": "64de6b20e549222d7456ec41b7e90d976a3186746092d94e04bb350c64dbcfd2", "time_to_first_visible_seconds": 26.839208, "time_to_last_output_seconds": 26.942446, "ttft_seconds": 26.839208, "wall_seconds": 26.942614 }, { "cache_salt_sha256": "e4309e295fbeb64d79c97adc4cf08618024d772d3f79d6ed878b89064d172961", "cached_prompt_tokens": 0, "completion_tokens": 8, "decode_seconds": 0.104047, "decode_tokens_per_second": 67.277, "effective_prefill_tokens_per_second": 2439.649, "finish_reason": "length", "measured_sse_events": 2, "output": " Each observation records behaviour, evidence, constraints", "output_characters": 58, "output_sha256": "3f91e56380da40aa484278ca6ad46278f1415f32fb168c9a22b4e5fb220e39c0", "prompt_tokens": 65536, "reasoning_characters": 0, "reasoning_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "requested_prompt_tokens": 65536, "response_tail_seconds": 0.000255, "run": 3, "stream_done_marker": true, "stream_sha256": "3f91e56380da40aa484278ca6ad46278f1415f32fb168c9a22b4e5fb220e39c0", "time_to_first_visible_seconds": 26.86288, "time_to_last_output_seconds": 26.966928, "ttft_seconds": 26.86288, "wall_seconds": 26.967183 } ], "ttft_seconds": { "maximum": 26.885, "median": 26.863, "minimum": 26.839, "p90": 26.885 } } }, "8192": { "cold": { "effective_prefill_tokens_per_second": { "maximum": 2583.761, "median": 2562.922, "minimum": 2519.777, "p90": 2583.761 }, "runs": [ { "cache_salt_sha256": "1181afb29f496b1e35a87d07eb49ab8b188c739e7079fb7c2fe81822bd8926d5", "cached_prompt_tokens": 0, "completion_tokens": 8, "decode_seconds": 0.072195, "decode_tokens_per_second": 96.96, "effective_prefill_tokens_per_second": 2519.777, "finish_reason": "length", "measured_sse_events": 2, "output": " tests, documentation, and release notes.", "output_characters": 41, "output_sha256": "052dc6176cb0a2e719a14647a717dfa4d0121a0bb7402d3dff7dd2217852cd4d", "prompt_tokens": 8192, "reasoning_characters": 0, "reasoning_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "requested_prompt_tokens": 8192, "response_tail_seconds": 0.000232, "run": 1, "stream_done_marker": true, "stream_sha256": "052dc6176cb0a2e719a14647a717dfa4d0121a0bb7402d3dff7dd2217852cd4d", "time_to_first_visible_seconds": 3.251081, "time_to_last_output_seconds": 3.323276, "ttft_seconds": 3.251081, "wall_seconds": 3.323508 }, { "cache_salt_sha256": "c9664648ee2eb87225bafd44a50e49bdc9cd054ab99992354a4921a0e55c9888", "cached_prompt_tokens": 0, "completion_tokens": 8, "decode_seconds": 0.074916, "decode_tokens_per_second": 93.438, "effective_prefill_tokens_per_second": 2562.922, "finish_reason": "length", "measured_sse_events": 2, "output": " documentation, and release notes. Each observation", "output_characters": 51, "output_sha256": "419fa9657b26154a535dc76436742d22a60ff64a8009c0fe929b11a35b6c39f5", "prompt_tokens": 8192, "reasoning_characters": 0, "reasoning_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "requested_prompt_tokens": 8192, "response_tail_seconds": 0.000289, "run": 2, "stream_done_marker": true, "stream_sha256": "419fa9657b26154a535dc76436742d22a60ff64a8009c0fe929b11a35b6c39f5", "time_to_first_visible_seconds": 3.196352, "time_to_last_output_seconds": 3.271268, "ttft_seconds": 3.196352, "wall_seconds": 3.271557 }, { "cache_salt_sha256": "99efbe3946d316f7d84d79c81cb54038436f7d4213964339726808d69ae93b9e", "cached_prompt_tokens": 0, "completion_tokens": 8, "decode_seconds": 0.073696, "decode_tokens_per_second": 94.984, "effective_prefill_tokens_per_second": 2583.761, "finish_reason": "length", "measured_sse_events": 2, "output": ", tests, documentation, and release notes", "output_characters": 41, "output_sha256": "41067174b9993d03b64fee1322cfb0c1ae84ccd92e980017d0765315051c4900", "prompt_tokens": 8192, "reasoning_characters": 0, "reasoning_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "requested_prompt_tokens": 8192, "response_tail_seconds": 0.000266, "run": 3, "stream_done_marker": true, "stream_sha256": "41067174b9993d03b64fee1322cfb0c1ae84ccd92e980017d0765315051c4900", "time_to_first_visible_seconds": 3.170572, "time_to_last_output_seconds": 3.244269, "ttft_seconds": 3.170572, "wall_seconds": 3.244534 } ], "ttft_seconds": { "maximum": 3.251, "median": 3.196, "minimum": 3.171, "p90": 3.251 } }, "warm_replay": { "effective_prefill_tokens_per_second": { "maximum": 2621.641, "median": 2606.23, "minimum": 2604.148, "p90": 2621.641 }, "runs": [ { "cache_salt_sha256": "1181afb29f496b1e35a87d07eb49ab8b188c739e7079fb7c2fe81822bd8926d5", "cached_prompt_tokens": 0, "completion_tokens": 8, "decode_seconds": 0.072193, "decode_tokens_per_second": 96.962, "effective_prefill_tokens_per_second": 2604.148, "finish_reason": "length", "measured_sse_events": 2, "output": " tests, documentation, and release notes.", "output_characters": 41, "output_sha256": "052dc6176cb0a2e719a14647a717dfa4d0121a0bb7402d3dff7dd2217852cd4d", "prompt_tokens": 8192, "reasoning_characters": 0, "reasoning_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "requested_prompt_tokens": 8192, "response_tail_seconds": 0.000161, "run": 1, "stream_done_marker": true, "stream_sha256": "052dc6176cb0a2e719a14647a717dfa4d0121a0bb7402d3dff7dd2217852cd4d", "time_to_first_visible_seconds": 3.14575, "time_to_last_output_seconds": 3.217944, "ttft_seconds": 3.14575, "wall_seconds": 3.218105 }, { "cache_salt_sha256": "c9664648ee2eb87225bafd44a50e49bdc9cd054ab99992354a4921a0e55c9888", "cached_prompt_tokens": 0, "completion_tokens": 8, "decode_seconds": 0.072179, "decode_tokens_per_second": 96.981, "effective_prefill_tokens_per_second": 2606.23, "finish_reason": "length", "measured_sse_events": 2, "output": " documentation, and release notes. Each observation", "output_characters": 51, "output_sha256": "419fa9657b26154a535dc76436742d22a60ff64a8009c0fe929b11a35b6c39f5", "prompt_tokens": 8192, "reasoning_characters": 0, "reasoning_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "requested_prompt_tokens": 8192, "response_tail_seconds": 0.000194, "run": 2, "stream_done_marker": true, "stream_sha256": "419fa9657b26154a535dc76436742d22a60ff64a8009c0fe929b11a35b6c39f5", "time_to_first_visible_seconds": 3.143237, "time_to_last_output_seconds": 3.215416, "ttft_seconds": 3.143237, "wall_seconds": 3.21561 }, { "cache_salt_sha256": "99efbe3946d316f7d84d79c81cb54038436f7d4213964339726808d69ae93b9e", "cached_prompt_tokens": 0, "completion_tokens": 8, "decode_seconds": 0.073679, "decode_tokens_per_second": 95.006, "effective_prefill_tokens_per_second": 2621.641, "finish_reason": "length", "measured_sse_events": 2, "output": ", tests, documentation, and release notes", "output_characters": 41, "output_sha256": "41067174b9993d03b64fee1322cfb0c1ae84ccd92e980017d0765315051c4900", "prompt_tokens": 8192, "reasoning_characters": 0, "reasoning_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "requested_prompt_tokens": 8192, "response_tail_seconds": 0.000237, "run": 3, "stream_done_marker": true, "stream_sha256": "41067174b9993d03b64fee1322cfb0c1ae84ccd92e980017d0765315051c4900", "time_to_first_visible_seconds": 3.124761, "time_to_last_output_seconds": 3.198441, "ttft_seconds": 3.124761, "wall_seconds": 3.198678 } ], "ttft_seconds": { "maximum": 3.146, "median": 3.143, "minimum": 3.125, "p90": 3.146 } } } }, "protocol": { "prompts_sha256": "0c3ac4015813af6a1b967659bd2289990fde895a5e43ae065b894d1e1f909ecb", "prompts_version": "1.0.0", "repository_dirty": false, "repository_revision": "770c3106729d5b73543afa2d9b498cebda6456f3", "repository_source": "git-worktree", "repository_source_sha256": "f5503bd3fe66d96d42992fedb6cf0b47f1e7ba560359d455a6e8fc1c471261d7", "version": "1.3.0" }, "run": { "appliance": { "competing_traffic": "none", "context_limit": 262144, "drafter": "checkpoint MTP head (Mia recipe DRAFTS=1 default)", "hardware": "1x NVIDIA DGX Spark (GB10, 128 GB unified memory)", "kv_cache_dtype": "fp8 (Mia recipe default)", "model": "azampatti/Qwen3.8-Flash-Next-125B-A5B-INT4-AutoRound", "model_revision": "1464274120d36a4d8fcaa934552334a7d83ce0fd", "negotiated_link_speed": "not applicable", "quantisation": "INT4 AutoRound (GPTQ int4 routed experts), FP8 n-gram table", "scheduler": "Mia recipe defaults (--parallel 8, shared KV pool) with VISION=0 and MAX_TOKENS=131072 in .env", "serving_engine": "TensorFold Zig engine (tensorfold-native) db281878 + Mia Qwen3.8-Flash-Next single-Spark kit (commit 364f21a)", "serving_image": "ghcr.io/miaai-lab/qwen3.8-flash-next-single-dgx-spark-tensorfold:zig-db28187-4144696f21e3 (image id sha256:12f9c5e0ce63f94aa7111736cedfcf4336195473ce0f917b2ad412880414005e)", "topology": "single Spark, TP1" }, "comparison_id": "q38-int4-1spark-xhigh", "finished_at": "2026-10-14T16:27:53.491258+00:00", "label": "q38-int4-base-xhigh", "model": "Qwen3.8-Flash-Next", "platform": "Linux-6.17.0-1026-nvidia-aarch64-with-glibc2.39", "python": "3.12.3", "rigmark_release_check": { "ahead_by": 5, "behind_by": 0, "dirty": false, "prerelease": true, "release_url": "https://github.com/alexellis/rigmark/releases/tag/v0.0.4", "revision": "770c3106729d5b73543afa2d9b498cebda6456f3", "status": "ahead", "tag": "v0.0.4" }, "started_at": "2026-10-14T15:56:33.530244+00:00" }, "schema": 1, "settings": { "concurrency": [ 1, 2, 4 ], "concurrency_runs": 3, "concurrency_tokens": 256, "concurrency_workload": "code", "decode_tokens": 32768, "extra_body": { "chat_template_kwargs": { "reasoning_effort": "xhigh" } }, "prefill_cache_isolation": "random-salt-per-pair", "prefill_depths": [ 8192, 32768, 65536 ], "prefill_runs": 3, "runs": 3, "seed": 20260905, "temperature": 0.0, "top_p": 1.0 } }