Skip to content
Closed
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 8 additions & 0 deletions opt_prime/examples/pd/http_matrix/.research/events.jsonl
Original file line number Diff line number Diff line change
Expand Up @@ -29,3 +29,11 @@
{"ts": "2026-10-24T05:13:10.885088+00:00", "task": "aicomp #265 / #267", "result": "KONI whole diagnostic264 responses pass;8K/128 native7.7847->whole9.3113tok/s; native/whole token parity6/8. Stream pending, different physical ATOM allocation.", "state": "active", "done": false, "review": true, "experiments": ["experiments/2026-10-24-stream265-atom-http-progress.md"], "runtime_commit": "33d6cb08ec806f693dbecfd7d630331a3bd1bb2f"}
{"ts": "2026-10-24T05:56:15.922056+00:00", "task": "aicomp #265 / #267", "result": "KONI all3 routes264 responses each audited; whole/stream8/8 exact.8K128 throughput7.7847/9.3113/10.8298tok/s. Native parity6/8; single batch and different physical card limits retained.", "state": "active", "done": false, "review": true, "experiments": ["experiments/2026-10-24-stream265-atom-http-progress.md"], "runtime_commit": "33d6cb08ec806f693dbecfd7d630331a3bd1bb2f"}
{"ts": "2026-10-30T04:24:12.257264+00:00", "task": "aicomp #201 / #265", "result": "ATOM short/long final-HTTP sweep:12 arms96conditions528responses audited;whole/stream16/16 references exact. Native/PD Qwen0.6/Qwen4=4/4,Gemma2/4(other2afterEOS),KONI2/4(preEOS). Eleven selfchecks pass;coordinator/cleanup0,fresh owned-Pod absence verified.", "state": "active", "done": false, "review": true, "experiments": ["experiments/2026-10-30-atom-edge201-final.md"], "runtime_scope": "historical stream265 B1 engine,final HTTP SHA256e3449979b88d03edb6d5bf0d402a708e5d98054fdf203e22b480f3d7e6fd759a"}
{"ts": "2026-10-30T05:14:21.740281+00:00", "task": "aicomp #267 / #201", "result": "HTTP backlog128 candidate:15CPUtests pass;Qwen4 whole/cacheoff24conditions344responses exact in one retry after originalreset. Remaining hardware verification pending.", "state": "active", "done": false, "review": true, "experiments": ["experiments/2026-10-30-http-backlog267.md"]}
{"ts": "2026-10-30T05:26:15.716445+00:00", "task": "aicomp #267 / #201", "result": "Qwen4 native ATOM cache-on retry:24 conditions,344 responses exact;13 auditor selfchecks pass;host listen drops +82 original/+0 retry. Wider hardware verification pending.", "state": "active", "done": false, "review": true, "experiments": ["experiments/2026-10-30-http-backlog267.md"]}
{"ts": "2026-10-30T05:36:34.408243+00:00", "task": "aicomp #267 / #201", "result": "Qwen4 RNGD-GPU whole KV paired cache off/on:48 conditions,688 responses exact. Across24 conditions throughput improves11,range -3.72% to +15.26%;single ordered pair. Full paired evidence retained;remaining hardware runs pending.", "state": "active", "done": false, "review": true, "experiments": ["experiments/2026-10-30-http-backlog267.md"]}
{"ts": "2026-10-30T05:48:16.366880+00:00", "task": "aicomp #267 / #201", "result": "Qwen4 GPU stream/off24conditions344responses exact;whole/stream references8/8equal. Same candidate,only stream flag differs. Single ordered pair all24 throughput increases(+4.33..30.38%);8K128c4 31.97->37.54tok/s. Remaining verification pending.", "state": "active", "done": false, "review": true, "experiments": ["experiments/2026-10-30-http-backlog267.md"]}
{"ts": "2026-10-30T06:03:17.316059+00:00", "task": "aicomp #267 / #201", "result": "Qwen4 GPU PD four configurations complete96conditions1376responses exact;all4reference sets equal,listen deltas0. Full grid provenance/timings retained;single ordered samples only. Gemma GPU and Qwen4 ATOM pending.", "state": "active", "done": false, "review": true, "experiments": ["experiments/2026-10-30-http-backlog267.md"]}
{"ts": "2026-10-30T06:38:39.539907+00:00", "task": "aicomp #267 / #201", "result": "Qwen4 ATOM whole/on24conditions344responses same-route exact,listen deltas0. Native/PD6/8refs equal;2preEOS differences also present in historical cacheoff. At8K128c4 native23.07vsPD15.04tok/s. Streaming pending.", "state": "active", "done": false, "review": true, "experiments": ["experiments/2026-10-30-http-backlog267.md"]}
{"ts": "2026-10-30T06:45:47.350695+00:00", "task": "aicomp #267 / #201", "result": "Gemma GPU PD four configurations96conditions1376responses exact,all4reference sets equal,listen deltas0. Full timing/provenance retained;streamcache gain10/24 only(-4.93..+2.58%),single ordered samples. Remaining model tests pending.", "state": "active", "done": false, "review": true, "experiments": ["experiments/2026-10-30-http-backlog267.md"]}
{"ts": "2026-10-30T06:50:26.376967+00:00", "task": "aicomp #267 / #201", "result": "Qwen4 ATOM native/whole/stream72conditions1032responses same-route exact;whole/stream8/8refs,nativePD6/8.8K128c4 native23.07/whole15.04/stream16.52tok/s. Owned cleanup and freshabsence verified. KONI GPU/TP2 pending.", "state": "active", "done": false, "review": true, "experiments": ["experiments/2026-10-30-http-backlog267.md"]}
18 changes: 18 additions & 0 deletions opt_prime/examples/pd/http_matrix/STATUS.md
Original file line number Diff line number Diff line change
Expand Up @@ -146,3 +146,21 @@ Four models, whole/stream, input128/8192, output32/128, c1/c4:64 conditions352 r
## ATOM short/long controls complete (2026-10-30)

All four models complete native/whole/stream at128/8192 input,32/128 output,c1/c4:96conditions528responses. Whole/stream16/16 reference pairs match. Native/PD equality is Qwen0.6/Qwen4=4/4,Gemma2/4(other2equal through EOS),KONI2/4(pre-EOS differences retained). Single-batch8K/128/c4 stream throughput44.24/14.47/43.87/10.44tok/s; one ATOM lane,extra RNGD for PD. Eleven independent auditor selfchecks pass;all exits0 and owned Pod absence verified. Cache and broader goal work remain active. See experiments/2026-10-30-atom-edge201-final.md.

## HTTP listener backlog candidate (#267)

Backlog128 handles a16connection accept-pause regression;15HTTPtests pass. Qwen4 RNGD-GPU whole/cacheoff retry24conditions344responses exact after originalc16 connection reset;listen drops+36before/+0retry(host-wide). One retry only;ATOM/remaining GPU verification pending. See experiments/2026-10-30-http-backlog267.md.

PR289 follow-up: Qwen4 native ATOM cache-on retry also passes 24 conditions/344 responses; listen drops +82 original/+0 retry, no OOM in original. Both GPU and ATOM have one successful retry; other routes/models remain pending.

PR289 paired-cache follow-up: Qwen4 RNGD-GPU whole KV paired cache off/on:48 conditions,688 responses exact. Across24 conditions throughput improves11,range -3.72% to +15.26%;single ordered pair. Full paired evidence retained;remaining hardware runs pending.

PR289 streamed-control follow-up: Qwen4 GPU stream/off24conditions344responses exact;whole/stream references8/8equal. Same candidate,only stream flag differs. Single ordered pair all24 throughput increases(+4.33..30.38%);8K128c4 31.97->37.54tok/s. Remaining verification pending.

PR289 Qwen4 GPU subset: Qwen4 GPU PD four configurations complete96conditions1376responses exact;all4reference sets equal,listen deltas0. Full grid provenance/timings retained;single ordered samples only. Gemma GPU and Qwen4 ATOM pending.

PR289 ATOM PD follow-up: Qwen4 ATOM whole/on24conditions344responses same-route exact,listen deltas0. Native/PD6/8refs equal;2preEOS differences also present in historical cacheoff. At8K128c4 native23.07vsPD15.04tok/s. Streaming pending.

PR289 Gemma GPU subset: Gemma GPU PD four configurations96conditions1376responses exact,all4reference sets equal,listen deltas0. Full timing/provenance retained;streamcache gain10/24 only(-4.93..+2.58%),single ordered samples. Remaining model tests pending.

PR289 Qwen4 ATOM subset: Qwen4 ATOM native/whole/stream72conditions1032responses same-route exact;whole/stream8/8refs,nativePD6/8.8K128c4 native23.07/whole15.04/stream16.52tok/s. Owned cleanup and freshabsence verified. KONI GPU/TP2 pending.
Loading