gahow/aituner
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Commit Graph

  • d23b69219b Add clean dash1 harness ablation runner Gahow Wang 2026-06-21 00:51:08 +08:00
  • 488fae7e63 Add tuning progress report for harness evaluation Gahow Wang 2026-06-21 00:48:21 +08:00
  • 426151bc9f Harness stop uses full state baseline Gahow Wang 2026-06-20 22:48:27 +08:00
  • a9d237bbfd Show effective flags in ablation trajectory Gahow Wang 2026-06-20 10:24:53 +08:00
  • 5257fbc1a2 Improve harness incumbent follow-up search Gahow Wang 2026-06-20 05:37:15 +08:00
  • b3156a382a Harness: gate gpu-mem-util/seqs-raise on 'no untested TP increase' (frontier-closed) Gahow Wang 2026-06-19 13:33:29 +08:00
  • 76cca89a43 Add harness-only dash1 driver to verify the gpu-mem-util fix recovers ~0.87 + stops Gahow Wang 2026-06-19 11:29:32 +08:00
  • 83162e7a64 Ablation: pin gpt-5.5 @ ai.gahow.org (chat.completions); re-read token per arm Gahow Wang 2026-06-19 11:27:47 +08:00
  • a3523f5601 Harness: explore gpu-memory-utilization (and raise max-num-seqs) before Stop-B Gahow Wang 2026-06-19 10:25:47 +08:00
  • 95c02d7dd9 Fig-18: chained driver for 2 extra naive runs (n=3 nondeterminism) Gahow Wang 2026-06-18 09:06:05 +08:00
  • a1b804f879 Ablation: search.high 0.25 -> 0.15 (skip wildly-infeasible top probes) Gahow Wang 2026-06-17 22:11:52 +08:00
  • 0c23285f39 Fig18 substrate: real output_length + criterion-A time_scale + Stop-A drain deadline Gahow Wang 2026-06-17 17:24:00 +08:00
  • 816765071f Complete harness-vs-naive ablation: harness 3x faster + stops; naive nondeterministic Gahow Wang 2026-06-17 13:03:26 +08:00
  • 97d2ddabb1 Ablation driver: force direct LLM connection (codex proxy is dash0-local) Gahow Wang 2026-06-17 10:05:44 +08:00
  • 8e58b4033d Note dash1 lacks LLM gateway access (naive-completion deferred to dash0) Gahow Wang 2026-06-17 09:55:39 +08:00
  • b779f6e56a Add dash1 naive-completion driver for the ablation Gahow Wang 2026-06-17 09:52:54 +08:00
  • e7d1b3ba01 Harness-vs-naive ablation result: harness steers to TP & converges; naive wanders Gahow Wang 2026-06-17 09:51:56 +08:00
  • 579dd86698 Ablation: --skip-baseline so loops climb from first proposal Gahow Wang 2026-06-16 20:59:46 +08:00
  • 37342a5749 Add chained harness-vs-naive ablation driver (sequential runs + DONE marker) Gahow Wang 2026-06-16 20:30:41 +08:00
  • 5965f4fbbc Ablation substrate: scale=0.5 + out=128 + 6 probes (TP1 measurable, tractable) Gahow Wang 2026-06-16 20:29:30 +08:00
  • a1cbab0e69 Document harness-vs-naive ablation: setup, substrate calibration, blocker Gahow Wang 2026-06-16 20:16:27 +08:00
  • 0794efa249 Reduce ablation probe budget to 3 per trial for tractability Gahow Wang 2026-06-16 20:01:19 +08:00
  • d975e57bb5 Scale ablation early-stop caps to the compressed window (scale=0.2) Gahow Wang 2026-06-16 19:49:57 +08:00
  • a16016a876 Add harness vs naive ablation configs (27b, scale=0.2 substrate) Gahow Wang 2026-06-16 19:31:23 +08:00
  • 07f5d92e1d Add consolidated two-stop summary doc Gahow Wang 2026-06-16 19:16:28 +08:00
  • f2ff0faebd Document Stop-B end-to-end on dense 27B: the improving climb + no-regression Gahow Wang 2026-06-16 18:07:00 +08:00
  • 4a64196a99 Add 27B Stop-B agentic-loop config (harness-driven, GPUs 2-7) Gahow Wang 2026-06-16 09:08:46 +08:00
  • b17b213575 Tear down the engine on SIGTERM instead of orphaning it Gahow Wang 2026-06-16 09:08:06 +08:00
  • 93ce339d61 Document 27B TP sweep: per-GPU rises sharply with TP (dense), opposite of MoE Gahow Wang 2026-06-16 01:54:40 +08:00
  • b1b74318f6 Pin 27B A/B to GPUs 2-7 (route around leaked GPU0/1 memory) Gahow Wang 2026-06-15 23:01:22 +08:00
  • 2fcaf80450 Wrap socket/timeout errors in HTTP client as HttpClientError Gahow Wang 2026-06-15 22:58:28 +08:00
  • 3541065675 Speed up 27B TP A/B: request_timeout 180s, search.high 0.125 Gahow Wang 2026-06-15 22:40:42 +08:00
  • 7678c7d5e8 Switch 27B TP A/B to length-aware TTFT SLO (4s + L_in/8k), widen search Gahow Wang 2026-06-15 20:35:23 +08:00
  • ed2bbe0323 Add linear_ms SLO rule (length-aware TTFT budget) Gahow Wang 2026-06-15 20:35:23 +08:00
  • 77af4ded2a Flag Stop-B e2e per-GPU trajectory as non-benchmark (saturation + smoke regime) Gahow Wang 2026-06-15 18:40:38 +08:00
  • 4f45b546a1 Add 27B TP A/B (deterministic ground-truth: does TP2 beat TP1 per-GPU) Gahow Wang 2026-06-15 18:39:54 +08:00
  • 90c3eb51c8 Document Stop-B end-to-end validation (Phase 5) Gahow Wang 2026-06-15 17:58:44 +08:00
  • 0b6beafeb8 Phase 5: widen search.high to 1.0 to force multi-iteration Stop-B convergence Gahow Wang 2026-06-15 17:12:32 +08:00
  • d4aff81691 Add Stop-B end-to-end config (agentic loop, Stop-A enabled) Gahow Wang 2026-06-15 17:05:39 +08:00
  • f31e9ccfd5 Record Stop-A boundary-guard A/B: correct verdict, ~38% replay saved Gahow Wang 2026-06-15 16:57:53 +08:00
  • 03e556f0ab Add Stop-A ON config (adaptive_stop enabled + boundary guard) for A/B Gahow Wang 2026-06-15 16:25:24 +08:00
  • dfc823f972 Add Stop-A SLO-boundary guard Gahow Wang 2026-06-15 16:25:24 +08:00
  • 9f52812753 Document Stop-A validation: calibration + GPU fidelity check Gahow Wang 2026-06-15 16:03:16 +08:00
  • 958739027a Fix Stop-A validation config: system vllm, cap max-model-len Gahow Wang 2026-06-15 15:22:48 +08:00
  • 0f57ee96a9 Drop LLM endpoint from Stop-A full-data config (baseline-only run) Gahow Wang 2026-06-15 15:19:46 +08:00
  • 43125f48cf Address review of two-stop branch Gahow Wang 2026-06-15 15:19:08 +08:00
  • 3af1d84ac0 Add Stop-A full-data validation config (real-time replay, no cap) Gahow Wang 2026-06-15 15:15:12 +08:00
  • 08e53fd897 Add Stop-A calibration script (CPU-only convergence curve) Gahow Wang 2026-06-15 15:10:02 +08:00
  • a8f903498d Add Stop-B authority: deterministic validator overrides LLM stop Gahow Wang 2026-06-15 14:45:14 +08:00
  • 51a9e4a007 Add Stop-A: offered-L-C-A convergence early-stop for replay Gahow Wang 2026-06-15 14:23:49 +08:00
  • 0f15bbc3f1 Make the offered-load axis session-coherent Gahow Wang 2026-06-15 14:16:06 +08:00
  • 6f8e3c95c1 Unify harness L-C-A on the canonical lca.WorkloadProfile Gahow Wang 2026-06-15 14:12:17 +08:00
  • 8b4116fad0 Add reference paper and qwen27b tpot25 16-iter notes Gahow Wang 2026-06-15 14:02:30 +08:00
  • 27d1c8fa92 Add L-C-A workload profile metric and CLI profile commands Gahow Wang 2026-06-15 14:02:24 +08:00
  • 984eb1f325 Document 8-GPU harness ablation results for qwen27b and qwen235b prefill Gahow Wang 2026-05-16 21:23:16 +08:00
  • d0c89dac48 Clean marked trial engine processes Gahow Wang 2026-05-16 15:51:04 +08:00
  • cf9b8b3f68 Clean vLLM process groups after parent exit Gahow Wang 2026-05-16 14:52:05 +08:00
  • 5a879a8592 Fix decode harness partial probe handling Gahow Wang 2026-05-16 14:18:07 +08:00
  • f18765b235 Document eight-GPU harness rerun Gahow Wang 2026-05-13 09:04:14 +08:00
  • 5c2958e6c1 Constrain harness topology by visible GPUs Gahow Wang 2026-05-13 01:25:31 +08:00
  • fb6d74a18c Document harness v2 rerun criteria Gahow Wang 2026-05-12 22:23:12 +08:00
  • e3ed775afd Fix harness SLO early-stop diagnosis Gahow Wang 2026-05-12 22:20:01 +08:00
  • ef359c8eea Document profile-driven harness run Gahow Wang 2026-05-12 21:40:19 +08:00
  • 17e9681ca0 Add profile-driven harness planner Gahow Wang 2026-05-12 21:28:44 +08:00
  • 63d6a111f4 Document profile-driven harness design Gahow Wang 2026-05-12 21:09:29 +08:00
  • 2d03b1cd4c Add SLO-driven topology frontier harness guard Gahow Wang 2026-05-12 21:00:49 +08:00
  • e1125475ae Minimize no-harness ablation prompt Gahow Wang 2026-05-12 09:42:53 +08:00
  • ae756600ce Support full-range and incumbent-floor search modes Gahow Wang 2026-05-11 12:58:46 +08:00
  • 8516cd88c0 Use full search range for every trial Gahow Wang 2026-05-11 12:50:22 +08:00
  • 14259fcec9 Measure lower-range performance for infeasible trials Gahow Wang 2026-05-10 14:30:34 +08:00
  • bf7c02e721 Clarify qwen27b raw per-iteration performance Gahow Wang 2026-05-10 14:24:10 +08:00
  • b0325ecfd9 Clarify qwen235b raw per-iteration performance Gahow Wang 2026-05-10 14:21:49 +08:00
  • 4cfd3757b6 Document qwen235b prefill harness ablation Gahow Wang 2026-05-10 13:05:49 +08:00
  • bdb08f6edc Handle missing streamed token metrics Gahow Wang 2026-05-10 02:40:00 +08:00
  • 307e2eb0e8 Document qwen27b harness ablation Gahow Wang 2026-05-10 01:12:21 +08:00
  • adc4351e5d Report latency stats for infeasible baseline Gahow Wang 2026-05-08 11:10:34 +08:00
  • eb137a0b62 Document TPOT40 baseline infeasible run Gahow Wang 2026-05-08 02:57:03 +08:00
  • f212673f44 Stop tuning when baseline is infeasible Gahow Wang 2026-05-08 01:07:36 +08:00
  • a7a5e9ad80 Make tune trial budget resumable Gahow Wang 2026-05-07 17:18:06 +08:00
  • 7263587cb6 clean: ci Gahow Wang 2026-05-06 22:56:53 +08:00
  • d7df1ebdac Add open source project metadata Gahow Wang 2026-05-06 21:18:21 +08:00
  • c1ff64381d Harden trial measurement accounting Gahow Wang 2026-05-06 21:18:09 +08:00
  • 871c4cfc02 Document qwen27b chat setup audit Gahow Wang 2026-05-06 20:32:09 +08:00
  • 98cd6dd81a Document qwen27b current config harness curve Gahow Wang 2026-05-06 18:00:43 +08:00
  • f653af09a8 Stop harness when feasible probe reaches search high Gahow Wang 2026-05-06 17:59:09 +08:00
  • 5d96689ea6 Make harness runtime refinement memory safe Gahow Wang 2026-05-06 17:37:31 +08:00
  • cf2e741550 Document high search rerun Gahow Wang 2026-05-06 03:19:51 +08:00
  • 0622e23817 Guide harness runtime refinement after TP Gahow Wang 2026-05-06 02:46:07 +08:00
  • 50067c926d Add harness guided first topology probe Gahow Wang 2026-05-06 02:28:46 +08:00
  • 915861b706 Document community vllm harness ablation Gahow Wang 2026-05-02 11:17:24 +08:00
  • 4c066c4e4e Stop harness when search high is saturated Gahow Wang 2026-05-02 11:04:59 +08:00
  • ccbf24ac47 Use time-compressed community vllm ablation Gahow Wang 2026-05-02 10:03:59 +08:00
  • d3d4c234f6 Bound community vllm ablation replay Gahow Wang 2026-05-02 09:58:56 +08:00
  • 4ef69cce78 Make harness stop conservative for ablation Gahow Wang 2026-05-02 09:47:16 +08:00
  • 664aeb49b2 Use local cache for qwen30b vllm runs Gahow Wang 2026-05-02 08:47:16 +08:00
  • 1880e859b5 Use vllm cu129 wheel on dash0 Gahow Wang 2026-05-02 08:28:23 +08:00
  • e215827503 Use uv auto torch backend for vllm 0.20 Gahow Wang 2026-05-02 08:21:27 +08:00
  • a7c9518ef6 Use local vllm venv for dash0 community run Gahow Wang 2026-05-02 08:17:04 +08:00
  • 1a3d628268 Add harness early stop ablation Gahow Wang 2026-05-02 08:08:14 +08:00
  • 6d3459c82d Document decode harness one-shot mechanism Gahow Wang 2026-05-02 06:25:06 +08:00