Commit Graph

455 Commits

Author SHA1 Message Date
c1c200b7cd Add decode batch-grid stability experiment 2026-07-23 16:55:55 +08:00
cb67ac8621 Reuse validated FlashInfer cache for TP2 smoke 2026-07-23 16:24:20 +08:00
fd859bc52c Use empty scp sync for TP2 fleet job 2026-07-23 16:12:46 +08:00
be523b1c07 Isolate TP2 smoke from dirty remote checkout 2026-07-23 16:11:13 +08:00
2c3220c2be Add TP2 prefill serving-path smoke experiment 2026-07-23 16:09:52 +08:00
9c1175a434 experiment: pin cu129 real pilot runtime 2026-07-20 19:02:40 +08:00
788270183d experiment: gate per-gpu sweep on control completion 2026-07-20 18:58:43 +08:00
e651ecc923 experiment: reuse predictors across load contracts 2026-07-20 18:51:31 +08:00
809ad9ffef experiment: add per-gpu workload control 2026-07-20 18:48:22 +08:00
a033a72195 analysis: summarize workload simulator regimes 2026-07-20 18:17:03 +08:00
dfe3f345d8 research: record workload sweep launch 2026-07-20 18:15:03 +08:00
157bf3668d fix: isolate simulator predictor cache 2026-07-20 18:11:33 +08:00
f727cbcf76 fix: sweep simulator families independently 2026-07-20 18:04:23 +08:00
f38e260639 fix: parse simulator group arguments 2026-07-20 18:01:38 +08:00
3453fbe522 experiment: add parallel workload simulator sweep 2026-07-20 17:54:20 +08:00
b302954dcb docs: correct Q30 MNS surface 2026-07-20 17:39:59 +08:00
ca999c4e49 fix: derive complete trace blocks from private artifact 2026-07-20 17:38:32 +08:00
75946d9d73 experiment: add workload regime taxonomy 2026-07-20 17:36:02 +08:00
39766141fb Decompose good/bad selection split across frozen surfaces
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-20 12:05:20 +08:00
18c0b25ae7 Conclude Qwen30 admission diagnosis 2026-07-19 20:56:36 +08:00
2970f74d67 Analyze Qwen30 admission amplification 2026-07-19 20:52:48 +08:00
a6c9beaa89 Diagnose Qwen30 Fixed-PD admission state 2026-07-19 20:41:20 +08:00
409d83a876 Record Qwen235 state diagnosis results 2026-07-19 19:18:57 +08:00
e602d41c80 Report exact state composition verdict 2026-07-19 19:16:20 +08:00
27b2daf7ce Distinguish proxy and exact state verdicts 2026-07-19 19:12:51 +08:00
3be6cd04ad Reschedule clean Qwen235 TP4 iteration state 2026-07-19 19:05:05 +08:00
51f2072d30 Use critical path totals for TP8 breakdown 2026-07-19 19:03:49 +08:00
6b80266aa7 Fix Qwen235 state reweighting 2026-07-19 19:01:36 +08:00
9c9479c313 Match Frontier to exact iteration composition 2026-07-19 19:00:18 +08:00
3349b23290 Analyze corrected Frontier op traces 2026-07-19 18:53:12 +08:00
d5c3c93577 Schedule Qwen235 exact iteration state runs 2026-07-19 18:51:15 +08:00
b49c072502 Add exact state observation modes 2026-07-19 18:49:06 +08:00
670eca176d Include mixed steps in Qwen235 state analysis 2026-07-19 18:39:36 +08:00
b6dfbfcad7 Analyze Qwen235 fixed PD simulator state 2026-07-19 18:37:18 +08:00
5927b6bfc3 Add Qwen235 state replay diagnosis 2026-07-19 18:31:47 +08:00
10567da523 Freeze Qwen235 ablation operator profiles 2026-07-19 17:11:26 +08:00
4ca295af0b Match FlashInfer TP8 workspace initialization 2026-07-19 16:54:24 +08:00
80ab724608 Run Qwen235 profiling from clean remote worktree 2026-07-19 16:47:12 +08:00
506e63a633 Add measured collective profile gate for Qwen235 2026-07-19 16:45:35 +08:00
f4a75aa8e4 Update prefix trace parser test contract 2026-07-19 15:35:17 +08:00
9c34650746 Merge simulator fidelity work into feature/sim 2026-07-19 15:34:28 +08:00
d17a2d4ab0 Merge fixed-PD pressure probe into feature/sim 2026-07-19 15:34:23 +08:00
3d3878c5aa Summarize Frontier selection regret 2026-07-19 15:31:16 +08:00
4c8d581a5b Track simulator fidelity experiment artifacts 2026-07-19 15:31:09 +08:00
e0ea7e9961 Support scp GPU fleet sync 2026-07-19 15:30:54 +08:00
fbf0f7c50b Set Qwen235 profile KV cache layout 2026-07-19 11:47:36 +08:00
81f3a5c76d Profile Qwen235 true mixed attention 2026-07-19 11:45:08 +08:00
71dbf80ed4 Match Qwen235 Frontier chunked prefill 2026-07-19 11:35:27 +08:00
5f48f7ec8b Avoid stale exact-trace HTTP connections 2026-07-19 10:58:38 +08:00
42e4ae3422 Avoid occupied Qwen235 trace ports 2026-07-19 08:10:29 +08:00