Commit Graph

9 Commits

Author SHA1 Message Date
wassname fdb4c77d6c Add reference-impl URLs to variant docstrings + V2 external review
- Fetch canonical reference impls for offline review:
  * peft_{lora,hra,delora,ia3}_layer.py + peft_lora_{dora,variants}.py
  * orig_pissa_init.py (MuLabPKU/PiSSA)
  * orig_hra_layer.py (DaShenZi721/HRA)
  * orig_delora.py (ExplainableML/DeLoRA author fork)
- Add reference-impl URLs to all 6 variant docstrings
- Document HRA gate=0 dead-grad issue and DoRA detach-omission in their docstrings
- Re-run external review (codex) with refs available -> docs/audit/variants_review_v2.md
  Major NEW findings vs paper-only review:
    * DeLoRA: scalar W.norm() should be per-input-channel norm(dim=0)
    * HRA: PEFT uses symmetric repeated-column init (no dead grad), not zero gate
    * IA3: FFN targets need input-side gating, not output, our up_proj advice wrong
    * All LoRA-family: cfg.dropout silently ignored (no-op)
    * DeLoRA: wnorm should be persistent buffer, not Parameter
  HRA and DeLoRA upgraded to BUGGY (from Partial)
2026-04-26 19:27:47 +08:00
wassname d0b4c52740 External review: per-variant audit + design notes
- Two acpx external reviews (codex + opencode):
  * docs/audit/variants_review.md: per-variant paper-vs-impl audit
  * docs/audit/design_review.md: peft EVA / baukit / antipasto3 vs lora-lite
  * docs/audit/SUMMARY.md: aggregate verdicts + 3 risks + 5 follow-ups
- docs/refs/: peft_eva.py, peft_eva_finetuning.py, baukit_nethook.py,
  antipasto3_svd_adapter.py for offline reference

Findings: LoRA clean; PiSSA/DoRA/IA3/HRA/DeLoRA have documented partial deviations.
Top risks: init/grad tradeoffs hidden by coarse tests; qwen probe lacks strict
identity tol; IA3 target placement untested.
2026-04-26 19:01:29 +08:00
wassname 0d929f93b3 feat(hra): add Householder Reflection Adaptation, hook-only/bnb-friendly + Qwen proof 2026-04-26 17:58:56 +08:00
wassname 43e620176c docs: record DoRA + IA3 Qwen-0.6B proof results (tasks 80, 81) 2026-04-26 17:54:54 +08:00
wassname 2abf616be6 feat(dora): add weight-decomposed LoRA variant for fp layers 2026-04-26 17:53:33 +08:00
wassname 699fde31bf feat: ia3 variant, real bnb 4bit/8bit smoke, dev guide split, user-only readme 2026-04-26 17:49:17 +08:00
wassname f2d9021511 ci: add publishable check workflow 2026-04-26 17:09:47 +08:00
wassname 69bf5f4e44 test: prove adapter training paths 2026-04-26 17:00:39 +08:00
wassname 4db5cee5a9 init 2026-04-26 14:10:20 +08:00