{"title":"All measured configurations","model":"Qwen3-30B-A3B-Instruct-2507 Q4_K_M","output_tokens":128,"note":"Record of every scored configuration. Phases, hardware and placement differ; compare within a phase.","unscored":[{"engine":"PXA Overdrive","outcome":"The switch did not activate an Overdrive recipe for these weights."},{"engine":"Strata v0.2.10","outcome":"Loader requires PLE tensors this model does not have."}],"configurations":[{"id":"v100-0982-parallel-prep","label":"expXa 0982 · parallel activation prep","engine":"expXa 0982fd3","hardware":"V100-SXM2 32GB (one GPU), 19.2-core CPU quota","phase":"2026-11-03 activation-prep study","settings":{"context":8192,"threads":8,"threads_batch":20,"cpu_expert_banks":40,"batch":1024,"ubatch":256,"kv":"q8_0 / q8_0"},"summary":[{"prompt_tokens":512,"samples":3,"prefill_tok_s":{"median":203.974,"min":199.394,"max":207.855},"decode_tok_s":{"median":37.281,"min":36.017,"max":38.703},"request_s":{"median":6.003,"min":5.772,"max":6.067}},{"prompt_tokens":4096,"samples":3,"prefill_tok_s":{"median":201.529,"min":175.579,"max":202.041},"decode_tok_s":{"median":32.957,"min":32.494,"max":34.05},"request_s":{"median":24.271,"min":24.163,"max":27.095}}],"runs":[{"repeat":0,"prompt_tokens":4096,"request_s":24.271,"prefill_tok_s":201.529,"decode_tok_s":32.494,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":0,"prompt_tokens":512,"request_s":6.003,"prefill_tok_s":199.394,"decode_tok_s":37.281,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":1,"prompt_tokens":4096,"request_s":24.163,"prefill_tok_s":202.041,"decode_tok_s":32.957,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":1,"prompt_tokens":512,"request_s":6.067,"prefill_tok_s":203.974,"decode_tok_s":36.017,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":2,"prompt_tokens":4096,"request_s":27.095,"prefill_tok_s":175.579,"decode_tok_s":34.05,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":2,"prompt_tokens":512,"request_s":5.772,"prefill_tok_s":207.855,"decode_tok_s":38.703,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"}]},{"id":"v100-0982-old-prep","label":"expXa 0982 · old prep path (ablation)","engine":"expXa 0982fd3","hardware":"V100-SXM2 32GB (one GPU), 19.2-core CPU quota","phase":"2026-11-03 activation-prep study","settings":{"context":8192,"threads":8,"threads_batch":20,"cpu_expert_banks":40,"batch":1024,"ubatch":256,"kv":"q8_0 / q8_0"},"summary":[{"prompt_tokens":512,"samples":3,"prefill_tok_s":{"median":190.344,"min":189.568,"max":197.826},"decode_tok_s":{"median":38.159,"min":37.505,"max":38.931},"request_s":{"median":6.004,"min":5.991,"max":6.046}},{"prompt_tokens":4096,"samples":3,"prefill_tok_s":{"median":189.38,"min":188.365,"max":191.263},"decode_tok_s":{"median":30.925,"min":29.077,"max":32.873},"request_s":{"median":25.825,"min":25.528,"max":25.894}}],"runs":[{"repeat":0,"prompt_tokens":4096,"request_s":25.894,"prefill_tok_s":188.365,"decode_tok_s":30.925,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":0,"prompt_tokens":512,"request_s":6.046,"prefill_tok_s":190.344,"decode_tok_s":38.159,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":1,"prompt_tokens":4096,"request_s":25.528,"prefill_tok_s":189.38,"decode_tok_s":32.873,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":1,"prompt_tokens":512,"request_s":6.004,"prefill_tok_s":197.826,"decode_tok_s":37.505,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":2,"prompt_tokens":4096,"request_s":25.825,"prefill_tok_s":191.263,"decode_tok_s":29.077,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":2,"prompt_tokens":512,"request_s":5.991,"prefill_tok_s":189.568,"decode_tok_s":38.931,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"}]},{"id":"v100-pxa4-prep-control","label":"PXA 4.0.1 · control for 0982","engine":"PXA 4.0.1 caf43bd939","hardware":"V100-SXM2 32GB (one GPU), 19.2-core CPU quota","phase":"2026-11-03 activation-prep study","settings":{"context":8192,"threads":8,"threads_batch":20,"cpu_expert_banks":40,"batch":1024,"ubatch":256,"kv":"q8_0 / q8_0"},"summary":[{"prompt_tokens":512,"samples":3,"prefill_tok_s":{"median":180.822,"min":152.688,"max":199.263},"decode_tok_s":{"median":31.232,"min":31.065,"max":38.33},"request_s":{"median":6.954,"min":5.911,"max":7.454}},{"prompt_tokens":4096,"samples":3,"prefill_tok_s":{"median":202.001,"min":182.11,"max":203.135},"decode_tok_s":{"median":30.406,"min":30.103,"max":34.313},"request_s":{"median":24.382,"min":24.014,"max":26.759}}],"runs":[{"repeat":0,"prompt_tokens":4096,"request_s":26.759,"prefill_tok_s":182.11,"decode_tok_s":30.103,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":0,"prompt_tokens":512,"request_s":6.954,"prefill_tok_s":180.822,"decode_tok_s":31.065,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":1,"prompt_tokens":4096,"request_s":24.014,"prefill_tok_s":202.001,"decode_tok_s":34.313,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":1,"prompt_tokens":512,"request_s":7.454,"prefill_tok_s":152.688,"decode_tok_s":31.232,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":2,"prompt_tokens":4096,"request_s":24.382,"prefill_tok_s":203.135,"decode_tok_s":30.406,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":2,"prompt_tokens":512,"request_s":5.911,"prefill_tok_s":199.263,"decode_tok_s":38.33,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"}]},{"id":"v100-a078-reference","label":"expXa a078 · reference for 0982","engine":"expXa a078474","hardware":"V100-SXM2 32GB (one GPU), 19.2-core CPU quota","phase":"2026-11-03 activation-prep study","settings":{"context":8192,"threads":8,"threads_batch":20,"cpu_expert_banks":40,"batch":1024,"ubatch":256,"kv":"q8_0 / q8_0"},"summary":[{"prompt_tokens":512,"samples":2,"prefill_tok_s":{"median":187.433,"min":183.492,"max":191.373},"decode_tok_s":{"median":37.75,"min":37.01,"max":38.489},"request_s":{"median":6.127,"min":6.004,"max":6.251}},{"prompt_tokens":4096,"samples":2,"prefill_tok_s":{"median":179.34,"min":175.356,"max":183.324},"decode_tok_s":{"median":32.462,"min":31.397,"max":33.528},"request_s":{"median":26.806,"min":26.169,"max":27.444}}],"runs":[{"repeat":0,"prompt_tokens":4096,"request_s":26.169,"prefill_tok_s":183.324,"decode_tok_s":33.528,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":0,"prompt_tokens":512,"request_s":6.251,"prefill_tok_s":183.492,"decode_tok_s":37.01,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":1,"prompt_tokens":4096,"request_s":27.444,"prefill_tok_s":175.356,"decode_tok_s":31.397,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":1,"prompt_tokens":512,"request_s":6.004,"prefill_tok_s":191.373,"decode_tok_s":38.489,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"}]},{"id":"v100-a078-cpu-experts","label":"expXa a078 · CPU expert path","engine":"expXa a078474","hardware":"V100-SXM2 32GB (one GPU), 19.2-core CPU quota","phase":"2026-11-03 CPU expert path study","settings":{"context":8192,"threads":8,"threads_batch":20,"cpu_expert_banks":40,"batch":1024,"ubatch":256,"kv":"q8_0 / q8_0"},"summary":[{"prompt_tokens":512,"samples":4,"prefill_tok_s":{"median":191.992,"min":182.137,"max":192.873},"decode_tok_s":{"median":35.601,"min":30.273,"max":38.194},"request_s":{"median":6.341,"min":6.008,"max":6.898}},{"prompt_tokens":4096,"samples":4,"prefill_tok_s":{"median":186.921,"min":185.763,"max":188.912},"decode_tok_s":{"median":33.183,"min":31.411,"max":33.407},"request_s":{"median":25.821,"min":25.68,"max":25.891}}],"runs":[{"repeat":0,"prompt_tokens":4096,"request_s":25.891,"prefill_tok_s":185.763,"decode_tok_s":33.407,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":0,"prompt_tokens":512,"request_s":6.291,"prefill_tok_s":182.137,"decode_tok_s":36.805,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":1,"prompt_tokens":4096,"request_s":25.875,"prefill_tok_s":186.074,"decode_tok_s":33.193,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":1,"prompt_tokens":512,"request_s":6.898,"prefill_tok_s":191.944,"decode_tok_s":30.273,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":2,"prompt_tokens":4096,"request_s":25.766,"prefill_tok_s":188.912,"decode_tok_s":31.411,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":2,"prompt_tokens":512,"request_s":6.008,"prefill_tok_s":192.873,"decode_tok_s":38.194,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":3,"prompt_tokens":4096,"request_s":25.68,"prefill_tok_s":187.767,"decode_tok_s":33.172,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":3,"prompt_tokens":512,"request_s":6.39,"prefill_tok_s":192.039,"decode_tok_s":34.397,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"}]},{"id":"v100-pxa4-a078-control","label":"PXA 4.0.1 · control for a078","engine":"PXA 4.0.1 caf43bd939","hardware":"V100-SXM2 32GB (one GPU), 19.2-core CPU quota","phase":"2026-11-03 CPU expert path study","settings":{"context":8192,"threads":8,"threads_batch":20,"cpu_expert_banks":40,"batch":1024,"ubatch":256,"kv":"q8_0 / q8_0"},"summary":[{"prompt_tokens":512,"samples":4,"prefill_tok_s":{"median":209.541,"min":202.007,"max":214.412},"decode_tok_s":{"median":37.617,"min":32.625,"max":38.132},"request_s":{"median":5.84,"min":5.766,"max":6.461}},{"prompt_tokens":4096,"samples":4,"prefill_tok_s":{"median":202.795,"min":186.154,"max":206.149},"decode_tok_s":{"median":33.65,"min":30.106,"max":33.885},"request_s":{"median":24.013,"min":23.655,"max":26.262}}],"runs":[{"repeat":0,"prompt_tokens":4096,"request_s":26.262,"prefill_tok_s":186.154,"decode_tok_s":30.106,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":0,"prompt_tokens":512,"request_s":5.896,"prefill_tok_s":206.41,"decode_tok_s":37.499,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":1,"prompt_tokens":4096,"request_s":24.202,"prefill_tok_s":200.987,"decode_tok_s":33.571,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":1,"prompt_tokens":512,"request_s":6.461,"prefill_tok_s":202.007,"decode_tok_s":32.625,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":2,"prompt_tokens":4096,"request_s":23.823,"prefill_tok_s":204.602,"decode_tok_s":33.729,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":2,"prompt_tokens":512,"request_s":5.766,"prefill_tok_s":212.672,"decode_tok_s":38.132,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":3,"prompt_tokens":4096,"request_s":23.655,"prefill_tok_s":206.149,"decode_tok_s":33.885,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":3,"prompt_tokens":512,"request_s":5.783,"prefill_tok_s":214.412,"decode_tok_s":37.735,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"}]},{"id":"v100-custom31-serving-refresh","label":"Custom PXA 3.1 · serving build, re-measured","engine":"Custom PXA 3.1 0a1fe56","hardware":"V100-SXM2 32GB (one GPU), 19.2-core CPU quota","phase":"2026-11-02 baseline","settings":{"context":8192,"threads":8,"threads_batch":20,"cpu_expert_banks":40,"batch":1024,"ubatch":256,"kv":"q8_0 / q8_0"},"summary":[{"prompt_tokens":512,"samples":3,"prefill_tok_s":{"median":123.236,"min":121.657,"max":124.009},"decode_tok_s":{"median":32.94,"min":29.221,"max":36.076},"request_s":{"median":8.1,"min":7.708,"max":8.515}},{"prompt_tokens":4096,"samples":3,"prefill_tok_s":{"median":125.609,"min":115.211,"max":127.373},"decode_tok_s":{"median":31.114,"min":29.524,"max":32.526},"request_s":{"median":36.559,"min":36.508,"max":39.681}}],"runs":[{"repeat":0,"prompt_tokens":4096,"request_s":36.559,"prefill_tok_s":125.609,"decode_tok_s":32.526,"output_sha256":"f458f8a4feb43fb229dec97f8d04fb74b3c74ce521bfb75fcf494e9c66aa207c"},{"repeat":0,"prompt_tokens":512,"request_s":7.708,"prefill_tok_s":123.236,"decode_tok_s":36.076,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":1,"prompt_tokens":4096,"request_s":39.681,"prefill_tok_s":115.211,"decode_tok_s":31.114,"output_sha256":"f458f8a4feb43fb229dec97f8d04fb74b3c74ce521bfb75fcf494e9c66aa207c"},{"repeat":1,"prompt_tokens":512,"request_s":8.1,"prefill_tok_s":121.657,"decode_tok_s":32.94,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":2,"prompt_tokens":4096,"request_s":36.508,"prefill_tok_s":127.373,"decode_tok_s":29.524,"output_sha256":"f458f8a4feb43fb229dec97f8d04fb74b3c74ce521bfb75fcf494e9c66aa207c"},{"repeat":2,"prompt_tokens":512,"request_s":8.515,"prefill_tok_s":124.009,"decode_tok_s":29.221,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"}]},{"id":"v100-custom31-clone","label":"Custom PXA 3.1 · isolated copy","engine":"Custom PXA 3.1 0a1fe56","hardware":"V100-SXM2 32GB (one GPU), 19.2-core CPU quota","phase":"2026-11-02 baseline","settings":{"context":8192,"threads":8,"threads_batch":20,"cpu_expert_banks":40,"batch":1024,"ubatch":256,"kv":"q8_0 / q8_0"},"summary":[{"prompt_tokens":512,"samples":3,"prefill_tok_s":{"median":108.575,"min":50.811,"max":109.777},"decode_tok_s":{"median":33.134,"min":30.229,"max":36.684},"request_s":{"median":8.953,"min":8.53,"max":13.57}},{"prompt_tokens":4096,"samples":3,"prefill_tok_s":{"median":128.128,"min":110.694,"max":135.318},"decode_tok_s":{"median":30.104,"min":26.729,"max":32.854},"request_s":{"median":36.766,"min":34.531,"max":40.918}}],"runs":[{"repeat":0,"prompt_tokens":4096,"request_s":36.766,"prefill_tok_s":128.128,"decode_tok_s":26.729,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":0,"prompt_tokens":512,"request_s":13.57,"prefill_tok_s":50.811,"decode_tok_s":36.684,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":1,"prompt_tokens":4096,"request_s":40.918,"prefill_tok_s":110.694,"decode_tok_s":32.854,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":1,"prompt_tokens":512,"request_s":8.953,"prefill_tok_s":108.575,"decode_tok_s":30.229,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":2,"prompt_tokens":4096,"request_s":34.531,"prefill_tok_s":135.318,"decode_tok_s":30.104,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":2,"prompt_tokens":512,"request_s":8.53,"prefill_tok_s":109.777,"decode_tok_s":33.134,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"}]},{"id":"v100-pxa4-baseline","label":"PXA 4.0.1 · same settings","engine":"PXA 4.0.1 caf43bd939","hardware":"V100-SXM2 32GB (one GPU), 19.2-core CPU quota","phase":"2026-11-02 baseline","settings":{"context":8192,"threads":8,"threads_batch":20,"cpu_expert_banks":40,"batch":1024,"ubatch":256,"kv":"q8_0 / q8_0"},"summary":[{"prompt_tokens":512,"samples":3,"prefill_tok_s":{"median":206.846,"min":119.847,"max":211.331},"decode_tok_s":{"median":37.416,"min":36.025,"max":38.952},"request_s":{"median":5.846,"min":5.764,"max":7.827}},{"prompt_tokens":4096,"samples":3,"prefill_tok_s":{"median":199.574,"min":173.912,"max":206.353},"decode_tok_s":{"median":29.302,"min":27.655,"max":34.319},"request_s":{"median":25.159,"min":23.59,"max":27.929}}],"runs":[{"repeat":0,"prompt_tokens":4096,"request_s":23.59,"prefill_tok_s":206.353,"decode_tok_s":34.319,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":0,"prompt_tokens":512,"request_s":7.827,"prefill_tok_s":119.847,"decode_tok_s":36.025,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":1,"prompt_tokens":4096,"request_s":27.929,"prefill_tok_s":173.912,"decode_tok_s":29.302,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":1,"prompt_tokens":512,"request_s":5.846,"prefill_tok_s":211.331,"decode_tok_s":37.416,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":2,"prompt_tokens":4096,"request_s":25.159,"prefill_tok_s":199.574,"decode_tok_s":27.655,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":2,"prompt_tokens":512,"request_s":5.764,"prefill_tok_s":206.846,"decode_tok_s":38.952,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"}]},{"id":"v100-208e21d-baseline","label":"expXa 208e21d · same settings","engine":"expXa 208e21d","hardware":"V100-SXM2 32GB (one GPU), 19.2-core CPU quota","phase":"2026-11-02 baseline","settings":{"context":8192,"threads":8,"threads_batch":20,"cpu_expert_banks":40,"batch":1024,"ubatch":256,"kv":"q8_0 / q8_0"},"summary":[{"prompt_tokens":512,"samples":3,"prefill_tok_s":{"median":79.851,"min":63.622,"max":91.808},"decode_tok_s":{"median":31.941,"min":31.165,"max":36.915},"request_s":{"median":10.422,"min":9.047,"max":12.158}},{"prompt_tokens":4096,"samples":3,"prefill_tok_s":{"median":90.731,"min":90.449,"max":109.951},"decode_tok_s":{"median":29.213,"min":28.261,"max":29.996},"request_s":{"median":49.421,"min":41.642,"max":49.827}}],"runs":[{"repeat":0,"prompt_tokens":4096,"request_s":49.827,"prefill_tok_s":90.449,"decode_tok_s":28.261,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":0,"prompt_tokens":512,"request_s":10.422,"prefill_tok_s":79.851,"decode_tok_s":31.941,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":1,"prompt_tokens":4096,"request_s":49.421,"prefill_tok_s":90.731,"decode_tok_s":29.996,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":1,"prompt_tokens":512,"request_s":12.158,"prefill_tok_s":63.622,"decode_tok_s":31.165,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":2,"prompt_tokens":4096,"request_s":41.642,"prefill_tok_s":109.951,"decode_tok_s":29.213,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":2,"prompt_tokens":512,"request_s":9.047,"prefill_tok_s":91.808,"decode_tok_s":36.915,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"}]},{"id":"v100-custom31-serving-first","label":"Custom PXA 3.1 · serving build, first pass","engine":"Custom PXA 3.1 0a1fe56","hardware":"V100-SXM2 32GB (one GPU), 19.2-core CPU quota","phase":"2026-11-02 baseline","settings":{"context":8192,"threads":8,"threads_batch":20,"cpu_expert_banks":40,"batch":1024,"ubatch":256,"kv":"q8_0 / q8_0"},"summary":[{"prompt_tokens":512,"samples":3,"prefill_tok_s":{"median":135.886,"min":134.026,"max":140.645},"decode_tok_s":{"median":36.062,"min":35.455,"max":36.398},"request_s":{"median":7.342,"min":7.194,"max":7.384}},{"prompt_tokens":4096,"samples":3,"prefill_tok_s":{"median":135.54,"min":135.193,"max":136.377},"decode_tok_s":{"median":32.662,"min":32.372,"max":33.073},"request_s":{"median":34.103,"min":33.965,"max":34.263}}],"runs":[{"repeat":0,"prompt_tokens":4096,"request_s":34.103,"prefill_tok_s":135.54,"decode_tok_s":33.073,"output_sha256":"f458f8a4feb43fb229dec97f8d04fb74b3c74ce521bfb75fcf494e9c66aa207c"},{"repeat":0,"prompt_tokens":512,"request_s":7.384,"prefill_tok_s":135.886,"decode_tok_s":35.455,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":1,"prompt_tokens":4096,"request_s":33.965,"prefill_tok_s":136.377,"decode_tok_s":32.662,"output_sha256":"f458f8a4feb43fb229dec97f8d04fb74b3c74ce521bfb75fcf494e9c66aa207c"},{"repeat":1,"prompt_tokens":512,"request_s":7.342,"prefill_tok_s":134.026,"decode_tok_s":36.398,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":2,"prompt_tokens":4096,"request_s":34.263,"prefill_tok_s":135.193,"decode_tok_s":32.372,"output_sha256":"f458f8a4feb43fb229dec97f8d04fb74b3c74ce521bfb75fcf494e9c66aa207c"},{"repeat":2,"prompt_tokens":512,"request_s":7.194,"prefill_tok_s":140.645,"decode_tok_s":36.062,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"}]},{"id":"titan-208e21d-40-banks","label":"expXa 208e21d · 40 CPU expert banks","engine":"expXa 208e21d (SM61)","hardware":"Titan Xp 12GB, 3.3-core CPU quota, 22.75 GiB RAM cap","phase":"2026-11-02 Titan Xp study","settings":{"context":8192,"threads":8,"threads_batch":20,"cpu_expert_banks":40,"batch":1024,"ubatch":256,"kv":"q8_0 / q8_0"},"summary":[{"prompt_tokens":512,"samples":3,"prefill_tok_s":{"median":11.204,"min":11.122,"max":11.333},"decode_tok_s":{"median":9.431,"min":8.914,"max":9.441},"request_s":{"median":59.609,"min":58.738,"max":60.058}},{"prompt_tokens":4096,"samples":3,"prefill_tok_s":{"median":11.286,"min":11.209,"max":11.324},"decode_tok_s":{"median":8.461,"min":8.234,"max":8.745},"request_s":{"median":378.469,"min":376.858,"max":380.064}}],"runs":[{"repeat":0,"prompt_tokens":4096,"request_s":376.858,"prefill_tok_s":11.324,"decode_tok_s":8.461,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":0,"prompt_tokens":512,"request_s":59.609,"prefill_tok_s":11.122,"decode_tok_s":9.431,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":1,"prompt_tokens":4096,"request_s":378.469,"prefill_tok_s":11.286,"decode_tok_s":8.234,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":1,"prompt_tokens":512,"request_s":58.738,"prefill_tok_s":11.333,"decode_tok_s":9.441,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":2,"prompt_tokens":4096,"request_s":380.064,"prefill_tok_s":11.209,"decode_tok_s":8.745,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":2,"prompt_tokens":512,"request_s":60.058,"prefill_tok_s":11.204,"decode_tok_s":8.914,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"}]},{"id":"titan-208e21d-19-banks","label":"expXa 208e21d · 19 CPU expert banks","engine":"expXa 208e21d (SM61)","hardware":"Titan Xp 12GB, 3.3-core CPU quota, 22.75 GiB RAM cap","phase":"2026-11-02 Titan Xp study","settings":{"context":8192,"threads":3,"threads_batch":4,"cpu_expert_banks":19,"batch":1024,"ubatch":256,"kv":"q8_0 / q8_0"},"summary":[{"prompt_tokens":512,"samples":3,"prefill_tok_s":{"median":51.575,"min":51.413,"max":51.575},"decode_tok_s":{"median":23.984,"min":23.612,"max":24.269},"request_s":{"median":15.298,"min":15.204,"max":15.351}},{"prompt_tokens":4096,"samples":3,"prefill_tok_s":{"median":51.226,"min":50.889,"max":51.276},"decode_tok_s":{"median":22.252,"min":22.208,"max":22.698},"request_s":{"median":85.654,"min":85.607,"max":86.248}}],"runs":[{"repeat":0,"prompt_tokens":4096,"request_s":85.654,"prefill_tok_s":51.276,"decode_tok_s":22.208,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":0,"prompt_tokens":512,"request_s":15.298,"prefill_tok_s":51.413,"decode_tok_s":23.984,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":1,"prompt_tokens":4096,"request_s":86.248,"prefill_tok_s":50.889,"decode_tok_s":22.252,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":1,"prompt_tokens":512,"request_s":15.351,"prefill_tok_s":51.575,"decode_tok_s":23.612,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":2,"prompt_tokens":4096,"request_s":85.607,"prefill_tok_s":51.226,"decode_tok_s":22.698,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":2,"prompt_tokens":512,"request_s":15.204,"prefill_tok_s":51.575,"decode_tok_s":24.269,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"}]},{"id":"titan-pxa4-19-banks","label":"PXA 4.0.1 · 19 CPU expert banks","engine":"PXA 4.0.1 caf43bd939 (SM61)","hardware":"Titan Xp 12GB, 3.3-core CPU quota, 22.75 GiB RAM cap","phase":"2026-11-02 Titan Xp study","settings":{"context":8192,"threads":3,"threads_batch":4,"cpu_expert_banks":19,"batch":1024,"ubatch":256,"kv":"q8_0 / q8_0"},"summary":[{"prompt_tokens":512,"samples":3,"prefill_tok_s":{"median":71.003,"min":70.896,"max":71.015},"decode_tok_s":{"median":25.449,"min":24.729,"max":25.5},"request_s":{"median":12.243,"min":12.232,"max":12.401}},{"prompt_tokens":4096,"samples":3,"prefill_tok_s":{"median":70.751,"min":70.683,"max":70.834},"decode_tok_s":{"median":22.748,"min":22.747,"max":23.063},"request_s":{"median":63.507,"min":63.46,"max":63.528}}],"runs":[{"repeat":0,"prompt_tokens":4096,"request_s":63.46,"prefill_tok_s":70.834,"decode_tok_s":22.747,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":0,"prompt_tokens":512,"request_s":12.232,"prefill_tok_s":71.015,"decode_tok_s":25.5,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":1,"prompt_tokens":4096,"request_s":63.507,"prefill_tok_s":70.683,"decode_tok_s":23.063,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":1,"prompt_tokens":512,"request_s":12.401,"prefill_tok_s":70.896,"decode_tok_s":24.729,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":2,"prompt_tokens":4096,"request_s":63.528,"prefill_tok_s":70.751,"decode_tok_s":22.748,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":2,"prompt_tokens":512,"request_s":12.243,"prefill_tok_s":71.003,"decode_tok_s":25.449,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"}]},{"id":"titan-208e21d-20-banks","label":"expXa 208e21d · 20 CPU expert banks","engine":"expXa 208e21d (SM61)","hardware":"Titan Xp 12GB, 3.3-core CPU quota, 22.75 GiB RAM cap","phase":"2026-11-02 Titan Xp study","settings":{"context":8192,"threads":3,"threads_batch":4,"cpu_expert_banks":20,"batch":1024,"ubatch":256,"kv":"q8_0 / q8_0"},"summary":[{"prompt_tokens":512,"samples":3,"prefill_tok_s":{"median":49.091,"min":48.5,"max":49.558},"decode_tok_s":{"median":24.253,"min":23.991,"max":24.668},"request_s":{"median":15.709,"min":15.523,"max":15.894}},{"prompt_tokens":4096,"samples":3,"prefill_tok_s":{"median":49.259,"min":49.02,"max":49.315},"decode_tok_s":{"median":22.407,"min":22.201,"max":22.571},"request_s":{"median":88.873,"min":88.831,"max":89.235}}],"runs":[{"repeat":0,"prompt_tokens":4096,"request_s":89.235,"prefill_tok_s":49.02,"decode_tok_s":22.571,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":0,"prompt_tokens":512,"request_s":15.709,"prefill_tok_s":49.091,"decode_tok_s":24.253,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":1,"prompt_tokens":4096,"request_s":88.873,"prefill_tok_s":49.259,"decode_tok_s":22.407,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":1,"prompt_tokens":512,"request_s":15.523,"prefill_tok_s":49.558,"decode_tok_s":24.668,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":2,"prompt_tokens":4096,"request_s":88.831,"prefill_tok_s":49.315,"decode_tok_s":22.201,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":2,"prompt_tokens":512,"request_s":15.894,"prefill_tok_s":48.5,"decode_tok_s":23.991,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"}]},{"id":"titan-pxa4-20-banks","label":"PXA 4.0.1 · 20 CPU expert banks","engine":"PXA 4.0.1 caf43bd939 (SM61)","hardware":"Titan Xp 12GB, 3.3-core CPU quota, 22.75 GiB RAM cap","phase":"2026-11-02 Titan Xp study","settings":{"context":8192,"threads":3,"threads_batch":4,"cpu_expert_banks":20,"batch":1024,"ubatch":256,"kv":"q8_0 / q8_0"},"summary":[{"prompt_tokens":512,"samples":3,"prefill_tok_s":{"median":67.924,"min":67.341,"max":67.931},"decode_tok_s":{"median":23.307,"min":21.977,"max":23.914},"request_s":{"median":13.032,"min":12.892,"max":13.429}},{"prompt_tokens":4096,"samples":3,"prefill_tok_s":{"median":67.558,"min":55.108,"max":67.69},"decode_tok_s":{"median":21.589,"min":21.566,"max":22.558},"request_s":{"median":66.455,"min":66.309,"max":80.326}}],"runs":[{"repeat":0,"prompt_tokens":4096,"request_s":80.326,"prefill_tok_s":55.108,"decode_tok_s":21.589,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":0,"prompt_tokens":512,"request_s":13.429,"prefill_tok_s":67.341,"decode_tok_s":21.977,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":1,"prompt_tokens":4096,"request_s":66.455,"prefill_tok_s":67.69,"decode_tok_s":21.566,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":1,"prompt_tokens":512,"request_s":12.892,"prefill_tok_s":67.931,"decode_tok_s":23.914,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"},{"repeat":2,"prompt_tokens":4096,"request_s":66.309,"prefill_tok_s":67.558,"decode_tok_s":22.558,"output_sha256":"42ea2d0b334833ccea0bf938eb2066ec8cd20bc0944b4130e8a786f227189a31"},{"repeat":2,"prompt_tokens":512,"request_s":13.032,"prefill_tok_s":67.924,"decode_tok_s":23.307,"output_sha256":"3eee8815aaaad9147470353202d912f19f92c766e3bf726ec6a6b63018cf87a1"}]}]}
