 EthanfelandClaude Opus 4.6
|
7959ce81fd
|
merge: integrate PR #1 correctness refactor into test branch
Merges srv1n's correctness refactor (PR #1) with conflict resolution
to preserve our recent setting renames and fixes while incorporating
their core improvements:
- Target-key grouping: analysis/merge operates on resolved model weights
instead of raw LoRA prefixes, fixing trainer-specific aliasing bugs
- Exact low-rank linear patches: lossless linear merges for weighted_sum
and weighted_average when no nonlinear ops are needed
- Excess conflict metric: sign disagreement beyond statistical expectation
(cosine-sim aware), replacing raw conflict ratio in decisions
- Subspace overlap: measures how much LoRA weight perturbations align
- Block smoothing: reduces noisy per-prefix mode flips
- Activation-aware calibration scoring for Phase 2 quality evaluation
- ACE-Step architecture support
- Soft LoKr/LoHa imports for graceful degradation
- RS-LoRA compensation re-enabled with safe heuristic
- 519-line test suite (19 tests, all passing)
Conflict resolution kept our:
- Setting renames (merge_refinement, patch_compression, strategy_set)
- Qwen-Image detection fix
- LoRACompatibilityAnalyzer node
- Conflict-mode masking in _apply_conflict_modes (PR's approach)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
2026-03-08 14:19:34 +01:00 |
|