Commit Graph
140 Commits
Author SHA1 Message Date
Joseph Magly 9e21621071 fix(colab): verify model access before loading the pipeline 2026-09-07 22:32:18 -04:00
Joseph Magly afe6f6465d Merge current main for CLI layer-selection validation 2026-09-07 17:17:06 -04:00
Joseph Magly b54bfe3955 style(test): remove trailing blank line from Optuna contracts 2026-09-07 16:26:56 -04:00
Joseph Magly 707e762904 fix(packaging): install and validate the Bayesian optimizer by default 2026-09-07 16:23:48 -04:00
Joseph Magly 40236a150e fix(bayesian): reset attempt state and persist every skipped outcome 2026-09-07 16:21:29 -04:00
Joseph Magly bbf06fd6c8 test(cli): verify layer selection reaches local and remote execution 2026-09-07 16:20:38 -04:00
Raja Mukerji 5fe46d2dca feat(cli): expose --layer-selection so layer strategy is reachable without the Python API 2026-09-07 09:43:39 -07:00
Raja Mukerji 2d31d64857 fix(bayesian): surface skipped optimization and clamp warm start to search space 2026-09-07 00:36:39 -07:00
Joseph Magly 985c9e9363 feat: add safe distributed checkpoint intake and preflight 2026-09-04 19:43:53 -04:00
Joseph Magly 131a4a8b55 docs: align fused offload claims with test contract 2026-09-02 14:32:29 -04:00
Pliny the YoungerandClaude Fable 5.1 89b63ce216 feat(abliteration): update offloaded fused MoE expert tensors in place (#115)
transformers 5 stores routed experts as fused 3D parameters
(experts.gate_up_proj / experts.down_proj) for Mixtral, Qwen3-MoE,
DeepSeek-V3, GLM-4 MoE, Llama 4, gpt-oss and OLMoE. When such a layer is
CPU- or disk-offloaded by Accelerate, EXCISE failed closed with
"offloaded fused expert tensors are not yet a supported surgery layout",
so every frontier MoE checkpoint large enough to need offload could not be
abliterated at all.

Route the three fused projections and the fused bias projection through the
existing LogicalParameterTransaction adapter: a meta-resident fused
parameter is resolved to a private copy of its authoritative backing value,
projected per expert, and committed atomically; the live parameter stays on
meta and quantized or unknown backing layouts still fail before mutation.
Thread offload_roots through the MoE dispatch call sites so parent-prefixed
Accelerate hooks resolve, and drop the preflight rejection of 3D meta
parameters (validate_offloaded_parameters already resolves them).

Tests: fused 3D, bias, granular and selective-inversion projections on
offloaded backing stores, bounded norm restoration, parent-prefixed hook
resolution, quantized fail-closed, commit-failure rollback, and a full
offline pipeline run on a disk-offloaded tiny Mixtral fixture that verifies
the saved checkpoint carries the updated experts with no meta tensors.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
2026-09-02 11:54:14 -04:00
Joseph Magly d6d4f5ba22 feat: evaluate saved Qwen3.8 checkpoints 2026-08-29 19:14:04 -04:00
Joseph Magly 95c0cd606d fix(eval): score full anchored completions 2026-08-29 19:00:51 -04:00
Joseph Magly 4b737cba64 feat(qwen38): add E03 coherence recovery candidate 2026-08-29 18:45:57 -04:00
Joseph Magly 941869d909 feat: add tune-only Qwen3.8 E02 candidate (#185) 2026-08-29 03:54:53 -04:00
Joseph Magly 02cafde149 fix: make supervised pipeline timeout configurable (#185) 2026-08-29 01:12:34 -04:00
Joseph Magly 9e149bf656 feat: add held-out Qwen3.8 experiment protocol (#185) 2026-08-29 00:17:41 -04:00
Joseph Magly e8a36d796c feat: add durable reconnect-safe experiment runs (#184) 2026-08-28 23:50:45 -04:00
Joseph Magly 41654bdcdf fix: make spectral evidence fail inconclusive (#183) 2026-08-28 23:25:14 -04:00
Joseph Magly 1f428f60cf fix: make KL optimization measured and reversible (#182) 2026-08-28 23:11:26 -04:00
Joseph Magly 683b39ba53 fix: surface bounded GPU admission failures 2026-08-28 10:42:41 -04:00
Joseph Magly 691a5da746 fix(qwen): recognize saved text derivatives 2026-08-27 20:17:09 -04:00
Joseph Magly 44f7842c9c fix(qwen): preserve validated runtime on checkpoint reload 2026-08-27 20:12:43 -04:00
Joseph Magly 827ba8e106 fix: validate Qwen hybrid runtime placement 2026-08-27 01:12:20 -04:00
Joseph Magly 270db39e27 fix(qwen): preserve checkpoint multi-EOS stops 2026-08-26 23:45:01 -04:00
Joseph Magly 2145a22d41 fix(gpu): permit bounded CUDA context residue 2026-08-26 23:21:03 -04:00
Joseph Magly f31d182456 test(gpu): cover benchmark lifecycle orchestration 2026-08-26 22:57:37 -04:00
Joseph Magly 6df26d07e7 test(gpu): keep benchmark contracts CPU-only 2026-08-26 22:51:13 -04:00
Joseph Magly 9ab258359e fix(gpu): admit benchmark loads through lifecycle broker 2026-08-26 22:48:59 -04:00
Joseph Magly 5aa11828a1 feat(surgery): add validated Qwen3.8 output-writer contract 2026-08-26 22:32:14 -04:00
Joseph Magly 1c6c7a47ca fix(validation): use chat contract for deterministic probes 2026-08-26 22:20:09 -04:00
Joseph Magly 31ceb7041c fix: gate GPU allocation on supervisor admission 2026-08-24 21:02:58 -04:00
Joseph Magly 4b677be612 feat: publish GPU model lifecycle events 2026-08-24 20:27:21 -04:00
Joseph Magly b8b7a12493 fix(pipeline): tolerate missing cancellation state 2026-08-24 14:36:39 -04:00
Joseph Magly b7969cfed0 test(cli): relax quality gates for random fixture 2026-08-24 13:57:23 -04:00
Joseph Magly ccba4f5d85 fix(surgery): guard unvalidated Qwen hybrid layouts 2026-08-24 13:54:31 -04:00
Joseph Magly d8d0a76231 fix(pipeline): fail closed on invalid output 2026-08-24 13:42:13 -04:00
Joseph Magly f354effc72 fix(ui): refresh GPU memory gauge periodically 2026-08-24 12:51:02 -04:00
Joseph Magly 8c4c68760d feat: make UI model loading configurable 2026-08-24 11:53:18 -04:00
Joseph Magly f6ee663e1b fix(device): fall back when psutil memory query fails 2026-08-23 21:48:06 -04:00
Joseph Magly 34218cf6f7 feat(models): add Qwen3.8 27B preset 2026-08-23 20:43:23 -04:00
Joseph Magly 3aaa7e563c test(ui): keep VRAM renderer dependency-light 2026-08-23 20:13:16 -04:00
Joseph Magly 5ce65e8193 fix(ui): show VRAM for every visible GPU 2026-08-23 20:06:08 -04:00
Joseph Magly cad100dd70 feat: refresh model catalog and shared-host guidance 2026-08-23 18:57:13 -04:00
Joseph Magly 6b6a0fe3d8 test(device): restore release branch coverage 2026-08-23 11:46:39 -04:00
Joseph Magly d52c7fc406 test(ui): classify lifecycle regression as conditional 2026-08-23 11:21:27 -04:00
Joseph Magly ad883cef0b feat: harden local runtime and verification 2026-08-23 11:16:10 -04:00
Joseph Magly beb4707401 ci: release tested source snapshots only 2026-08-22 11:28:30 -04:00
Joseph Magly c546d30e9e test: support TOML parsing on Python 3.10 2026-08-22 11:04:56 -04:00
Joseph Magly 7b7109feb5 ci: bind release evidence to artifact hashes 2026-08-22 10:59:08 -04:00