{"title":"Runtime tile-mode control","engine":"expXa 30cec2f","model":"Qwen3-30B-A3B-Instruct-2507 Q4_K_M","hardware":"Xeon Gold 6248 (AVX-512 VNNI), V100-SXM2 32GB, 19.2-core CPU quota","arms":{"0,0":"per-row tiles","1,0":"hoisted activation columns","1,1":"hoisted + VNNI"},"configuration":{"cpu_expert_layers":48,"generation_threads":8,"batch_threads":20,"context":8192,"kv":"Q8_0","speculation":"none","process_nice":10},"functional_checks":["18 CPU parity cases including native VNNI","VNNI debugger breakpoint","Exact GPU matrix checks on both V100s","All completion routes streaming/nonstreaming","Busy and deferred switches rejected without generation change","Hot-swap GET/POST rejected with 501","All real-model outputs match within each prompt length"],"summary":[{"prompt_tokens":512,"outputs_identical":true,"groups":[{"arm":"0,0","n":4,"prefill_median":178.594,"decode_median":28.466,"prefill_throttled_ms_median":2468.4},{"arm":"1,0","n":4,"prefill_median":200.959,"decode_median":30.563,"prefill_throttled_ms_median":2179.0},{"arm":"1,1","n":4,"prefill_median":234.963,"decode_median":30.583,"prefill_throttled_ms_median":1852.1}]},{"prompt_tokens":4096,"outputs_identical":true,"groups":[{"arm":"0,0","n":2,"prefill_median":184.278,"decode_median":25.821,"prefill_throttled_ms_median":14521.5},{"arm":"1,0","n":2,"prefill_median":194.496,"decode_median":26.027,"prefill_throttled_ms_median":14128.4},{"arm":"1,1","n":2,"prefill_median":229.761,"decode_median":25.572,"prefill_throttled_ms_median":10237.2}]}],"runs":[{"arm":"1,0","prompt_tokens":512,"pair":0,"position":0,"warmup":true,"prefill_tok_s":191.124,"decode_tok_s":30.261,"request_s":6.921,"prefill_throttled_ms":2443.8,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"1,0","prompt_tokens":512,"pair":0,"position":0,"warmup":false,"prefill_tok_s":193.209,"decode_tok_s":30.467,"request_s":6.866,"prefill_throttled_ms":2394.8,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"1,1","prompt_tokens":512,"pair":0,"position":1,"warmup":true,"prefill_tok_s":223.014,"decode_tok_s":30.334,"request_s":6.529,"prefill_throttled_ms":1800.7,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"1,1","prompt_tokens":512,"pair":0,"position":1,"warmup":false,"prefill_tok_s":222.569,"decode_tok_s":30.559,"request_s":6.502,"prefill_throttled_ms":2005.8,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"0,0","prompt_tokens":512,"pair":0,"position":2,"warmup":true,"prefill_tok_s":196.973,"decode_tok_s":27.933,"request_s":7.194,"prefill_throttled_ms":2146.1,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"0,0","prompt_tokens":512,"pair":0,"position":2,"warmup":false,"prefill_tok_s":172.579,"decode_tok_s":29.143,"request_s":7.373,"prefill_throttled_ms":2747.8,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"1,0","prompt_tokens":512,"pair":1,"position":0,"warmup":true,"prefill_tok_s":208.312,"decode_tok_s":30.446,"request_s":6.675,"prefill_throttled_ms":1956.0,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"1,0","prompt_tokens":512,"pair":1,"position":0,"warmup":false,"prefill_tok_s":208.368,"decode_tok_s":30.691,"request_s":6.641,"prefill_throttled_ms":2019.1,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"1,1","prompt_tokens":512,"pair":1,"position":1,"warmup":true,"prefill_tok_s":238.638,"decode_tok_s":30.829,"request_s":6.31,"prefill_throttled_ms":1756.2,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"1,1","prompt_tokens":512,"pair":1,"position":1,"warmup":false,"prefill_tok_s":235.73,"decode_tok_s":30.823,"request_s":6.337,"prefill_throttled_ms":1933.3,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"0,0","prompt_tokens":512,"pair":1,"position":2,"warmup":true,"prefill_tok_s":198.062,"decode_tok_s":26.777,"request_s":7.378,"prefill_throttled_ms":2184.7,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"0,0","prompt_tokens":512,"pair":1,"position":2,"warmup":false,"prefill_tok_s":175.939,"decode_tok_s":27.788,"request_s":7.53,"prefill_throttled_ms":2492.6,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"1,0","prompt_tokens":512,"pair":2,"position":0,"warmup":true,"prefill_tok_s":186.097,"decode_tok_s":25.921,"request_s":7.702,"prefill_throttled_ms":2378.0,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"1,0","prompt_tokens":512,"pair":2,"position":0,"warmup":false,"prefill_tok_s":193.55,"decode_tok_s":30.471,"request_s":6.859,"prefill_throttled_ms":2274.4,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"1,1","prompt_tokens":512,"pair":2,"position":1,"warmup":true,"prefill_tok_s":238.121,"decode_tok_s":30.587,"request_s":6.341,"prefill_throttled_ms":1852.3,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"1,1","prompt_tokens":512,"pair":2,"position":1,"warmup":false,"prefill_tok_s":239.707,"decode_tok_s":30.566,"request_s":6.337,"prefill_throttled_ms":1770.8,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"0,0","prompt_tokens":512,"pair":2,"position":2,"warmup":true,"prefill_tok_s":198.433,"decode_tok_s":30.538,"request_s":6.784,"prefill_throttled_ms":2272.3,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"0,0","prompt_tokens":512,"pair":2,"position":2,"warmup":false,"prefill_tok_s":181.25,"decode_tok_s":26.149,"request_s":7.733,"prefill_throttled_ms":2444.1,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"1,1","prompt_tokens":512,"pair":3,"position":0,"warmup":true,"prefill_tok_s":196.369,"decode_tok_s":23.921,"request_s":7.965,"prefill_throttled_ms":2163.4,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"1,1","prompt_tokens":512,"pair":3,"position":0,"warmup":false,"prefill_tok_s":234.197,"decode_tok_s":30.601,"request_s":6.383,"prefill_throttled_ms":1762.3,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"1,0","prompt_tokens":512,"pair":3,"position":1,"warmup":true,"prefill_tok_s":206.415,"decode_tok_s":30.702,"request_s":6.662,"prefill_throttled_ms":2178.7,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"1,0","prompt_tokens":512,"pair":3,"position":1,"warmup":false,"prefill_tok_s":209.1,"decode_tok_s":30.655,"request_s":6.629,"prefill_throttled_ms":2083.6,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"0,0","prompt_tokens":512,"pair":3,"position":2,"warmup":true,"prefill_tok_s":199.116,"decode_tok_s":30.601,"request_s":6.759,"prefill_throttled_ms":2211.5,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"0,0","prompt_tokens":512,"pair":3,"position":2,"warmup":false,"prefill_tok_s":195.384,"decode_tok_s":29.217,"request_s":7.007,"prefill_throttled_ms":2178.9,"output_sha256":"3aae83ae236c41a886a4e88f5106b166d535c9bb2dd4cdd9611eb6209f01c957"},{"arm":"1,0","prompt_tokens":4096,"pair":0,"position":0,"warmup":true,"prefill_tok_s":178.574,"decode_tok_s":28.041,"request_s":27.512,"prefill_throttled_ms":15497.6,"output_sha256":"655a8dc65683f84a9ef8dbff90ca473366bdb82cc68750cfe25ce8517b822c2d"},{"arm":"1,0","prompt_tokens":4096,"pair":0,"position":0,"warmup":false,"prefill_tok_s":194.487,"decode_tok_s":24.081,"request_s":26.396,"prefill_throttled_ms":14484.2,"output_sha256":"655a8dc65683f84a9ef8dbff90ca473366bdb82cc68750cfe25ce8517b822c2d"},{"arm":"0,0","prompt_tokens":4096,"pair":0,"position":1,"warmup":true,"prefill_tok_s":182.123,"decode_tok_s":26.749,"request_s":27.296,"prefill_throttled_ms":13722.0,"output_sha256":"655a8dc65683f84a9ef8dbff90ca473366bdb82cc68750cfe25ce8517b822c2d"},{"arm":"0,0","prompt_tokens":4096,"pair":0,"position":1,"warmup":false,"prefill_tok_s":176.747,"decode_tok_s":23.791,"request_s":28.574,"prefill_throttled_ms":16296.1,"output_sha256":"655a8dc65683f84a9ef8dbff90ca473366bdb82cc68750cfe25ce8517b822c2d"},{"arm":"1,1","prompt_tokens":4096,"pair":0,"position":2,"warmup":true,"prefill_tok_s":218.713,"decode_tok_s":27.867,"request_s":23.339,"prefill_throttled_ms":12488.4,"output_sha256":"655a8dc65683f84a9ef8dbff90ca473366bdb82cc68750cfe25ce8517b822c2d"},{"arm":"1,1","prompt_tokens":4096,"pair":0,"position":2,"warmup":false,"prefill_tok_s":231.847,"decode_tok_s":27.826,"request_s":22.279,"prefill_throttled_ms":9665.7,"output_sha256":"655a8dc65683f84a9ef8dbff90ca473366bdb82cc68750cfe25ce8517b822c2d"},{"arm":"1,1","prompt_tokens":4096,"pair":1,"position":0,"warmup":true,"prefill_tok_s":223.142,"decode_tok_s":27.966,"request_s":22.952,"prefill_throttled_ms":10736.4,"output_sha256":"655a8dc65683f84a9ef8dbff90ca473366bdb82cc68750cfe25ce8517b822c2d"},{"arm":"1,1","prompt_tokens":4096,"pair":1,"position":0,"warmup":false,"prefill_tok_s":227.675,"decode_tok_s":23.318,"request_s":23.499,"prefill_throttled_ms":10808.7,"output_sha256":"655a8dc65683f84a9ef8dbff90ca473366bdb82cc68750cfe25ce8517b822c2d"},{"arm":"1,0","prompt_tokens":4096,"pair":1,"position":1,"warmup":true,"prefill_tok_s":175.651,"decode_tok_s":25.57,"request_s":28.341,"prefill_throttled_ms":17808.0,"output_sha256":"655a8dc65683f84a9ef8dbff90ca473366bdb82cc68750cfe25ce8517b822c2d"},{"arm":"1,0","prompt_tokens":4096,"pair":1,"position":1,"warmup":false,"prefill_tok_s":194.505,"decode_tok_s":27.973,"request_s":25.645,"prefill_throttled_ms":13772.6,"output_sha256":"655a8dc65683f84a9ef8dbff90ca473366bdb82cc68750cfe25ce8517b822c2d"},{"arm":"0,0","prompt_tokens":4096,"pair":1,"position":2,"warmup":true,"prefill_tok_s":176.517,"decode_tok_s":23.948,"request_s":28.561,"prefill_throttled_ms":17077.1,"output_sha256":"655a8dc65683f84a9ef8dbff90ca473366bdb82cc68750cfe25ce8517b822c2d"},{"arm":"0,0","prompt_tokens":4096,"pair":1,"position":2,"warmup":false,"prefill_tok_s":191.81,"decode_tok_s":27.851,"request_s":25.967,"prefill_throttled_ms":12746.9,"output_sha256":"655a8dc65683f84a9ef8dbff90ca473366bdb82cc68750cfe25ce8517b822c2d"}],"excluded_attempts":[{"reason":"A sampled thread exited mid-read (ESRCH); the whole attempt was excluded and restarted after a harness fix.","runs":3}]}
