SLERP merge silently returns only the second adapter: state_a aliases base_model storage (CR-A13-001) #407

Wave 1 · theme D1-5: Training, merging and export correctness · highest severity High · 1 finding(s) · effort small

Summary

Operator merges a code adapter and a support adapter with slerp 0.5/0.5, ships the 'merged' model; it is just the second adapter, the first specialist's behaviour is silently lost, and README/docs advertise SLERP as a working method.

Primary finding

CR-A13-001 — SLERP merge silently returns only the second adapter: state_a aliases base_model storage

severity High · reachability default · intent contract-violation · confidence confirmed · independently confirmed · reproduced end to end

What happens. _slerp_merge stores state_a = model_a.merge_and_unload().state_dict() without cloning (forgelm/merging.py:198-199). The state_dict tensors share storage with base_model's live parameters, so base_model.load_state_dict(base_state) (merging.py:203) copies the base back into state_a in place, and merging adapter B (merging.py:205-206) writes base+dB into the same storage. state_a and state_b are therefore identical, and slerp(v, v) takes the near-parallel branch and returns base+B whatever the weights. MergeResult still reports success with method 'slerp'. The SLERP tests mock PEFT with independent dicts, so they cannot see the aliasing.

Where.

• forgelm/merging.py:196-206 (state_a not cloned; base_state is the only clone)
• forgelm/merging.py:199
• forgelm/merging.py:203 (confirmed)
• forgelm/merging.py:209-232
• forgelm/merging.py:198-207
• forgelm/merging.py:209-235 (confirmed)
• tests/test_merging_algos.py:590-640 (TestSlerpMergeExercisesRealFunction mocks peft and base.state_dict with independent dicts)
• forgelm/merging.py:196-207
• forgelm/merging.py:198
• forgelm/merging.py:206

Evidence.

• Two real LoRA adapters (sft, dpo) merged with slerp weights 0.7/0.3: for all q_proj/v_proj tensors max|out-base-dpo_delta| < 1e-8 while |out-base-sft_delta| ~1e-4 (sl2.py). Entry order swapped (dpo first, sft second): output equals sft delta, not dpo (sl3.py). Linear and TIES recomputed independently and match (err 7.5e-9 and 0). Existing slerp tests mock the PEFT load (tests/test_merging_algos.py:592-640) and cannot see aliasing.
• slerp_mech.py (tiny-qwen + env/artifacts/sft as A, env/artifacts/dpo as B; torch 2.12.1, transformers 5.13.0, peft 0.19.1): state_a[k].data_ptr()==live param ptr -> True; after load_state_dict(base_state) state_a[k]==base -> True; after B merge torch.equal(state_a[k], state_b[k]) -> True, same storage True. slerp_repro.py: merge_peft_adapters(method='slerp') success=True; for weights (1,1) and (0.9,0.1) the output's max |diff| on the LoRA-touched q_proj is 7.45e-09 to B and 1.018e-04 to A, whil…
• m3.py with tiny-qwen and the real distinct fixture adapters artifacts/sft (A) and artifacts/dpo (B), q_proj weight of layer 0: A and B differ by 1.0e-4 max. For weights 1:1, 3:1 and 1:3 the SLERP result differs from the B-merged model by at most 7.5e-9 (float noise) and from the A-merged model by 1.0e-4; the expected lerp-equivalent differs from the result by 5.1e-5, 7.6e-5, 2.5e-5. The result is identical for all three weightings.

Recommended fix. Clone state_a (and state_b) immediately after merge_and_unload, e.g. {k: v.clone() for k, v in ...state_dict().items()}; add a non-mocked tiny-adapter test asserting the result differs from each input and respects weights.

Acceptance criteria

• A regression test reproduces the failure scenario and asserts the corrected behaviour.
• Documentation listed above is corrected in English and in its Turkish mirror in the same PR.
• Every finding in this issue is resolved or explicitly closed with a reason.

_{Found by the October 2026 ForgeLM code and documentation review against commit cab2947. Code links point at that commit; line numbers may have moved since. Finding IDs (CR-…, DOC-…, TECH-…) are stable references for this backlog.}