Giorgio Gilestro
8da0dac007
Adds the union-preserving recombination operator that llm_merge lacked (E8's max,
not mean): keep each specialist LoRA intact and SELECT the right one per prompt
(MoE router: oracle, or training-free nearest-centroid over base embeddings) or
per module (max_merge = winner-take-all by delta norm). src/llm/moe.py, kind
llm_moe, reuses the cached specialists.
Result — a clean regime boundary for "merge, don't average":
- 0.5B: union wins. Routing 0.74 / worst-family 0.43 > soup 0.64 / 0.26, with no
dilution (recovers each specialist's own-family peak). E8's max > mean in real
weights, because at a weak base averaging dilutes.
- 7B (Imperial CX3, L40S, 9 min): the ordering INVERTS. Fusion wins — soup 0.87 >
routing 0.84 > max_merge 0.78. Routing is capped at the best parent per family;
fusion blends and, given a capable base, COMPOSES beyond any parent (soup lists
0.62 > spec 0.57). Selection can't synthesise better than its best component;
averaging-that-composes can.
So "merge, don't average" (E4/E8) is a weak-parent / small-model law, not
universal: union wins under dilution, fusion wins under composition. Refines E8
(its additive-landscape max>mean assumed no compositional headroom). The operator
to want is fusion-that-composes + offspring selection = the directed-sex ideal
(E10) — the natural next experiment.
Honest riders: the learned router is trivially perfect (lexically-distinct
families), and router-free max_merge is the weakest union (not input-adaptive).
+2 router unit tests (127 green). Results in results/llm_moe{,_hpc}/ (parquet
gitignored per the reproducibility contract).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
3.8 KiB
llm_moe_hpc — union vs fusion recombination at scale (7B, Imperial CX3): the regime flips
Claim tested. The 0.5B llm_moe found that a union operator (route/select, never average) beats
fusion (soup/ties, average the deltas), because at a weak base averaging dilutes. But
llm_merge_hpc showed that at 7B fusion stops diluting and starts composing (soup 0.87 beat every
specialist). So the sharp question: does routing still beat fusion once the base is capable — or does
a strong base invert the ordering? Run on one L40S (46 GB) GPU of Imperial's CX3 HPC, 9 min
walltime, reusing the cached 7B specialists.
Setup. Base Qwen2.5-7B-Instruct, the three cached LoRA specialists from llm_merge_hpc
(lists/strings/arith, exact-match verifier), 200 test tasks/family, seed 1. Same five operators
as the 0.5B run: fusion (soup, ties) vs union (route:oracle, route:learned, max-merge).
Results (accuracy)
| operator | lists | strings | arith | overall | worst-family | router |
|---|---|---|---|---|---|---|
| base | 0.46 | 0.69 | 1.00 | 0.71 | 0.46 | — |
| best specialist (lists) | 0.57 | 0.74 | 1.00 | 0.77 | 0.57 | — |
| fuse: soup | 0.62 | 1.00 | 1.00 | 0.87 | 0.62 | — |
| fuse: ties | 0.62 | 1.00 | 0.99 | 0.87 | 0.62 | — |
| route: oracle | 0.57 | 0.97 | 0.96 | 0.84 | 0.57 | 1.00 |
| route: learned | 0.57 | 0.97 | 0.96 | 0.84 | 0.57 | 1.00 |
| max-merge | 0.48 | 0.90 | 0.96 | 0.78 | 0.48 | — |
The finding: "merge, don't average" is regime-dependent, and inverts at scale
- The ordering flips. At 0.5B, union > fusion (routing 0.74 > soup 0.64). At 7B, fusion > union (soup 0.87 > routing 0.84 > max-merge 0.78). The exact opposite winner.
- Why: routing is capped at the best parent; fusion can exceed it. Routing selects one intact specialist, so per family it can only reach that specialist's own score (lists 0.57 = spec_lists, strings 0.97 = spec_strings). Fusion blends the deltas — and at a capable base the blend composes beyond any parent: soup scores lists 0.62 (> spec_lists 0.57) and strings 1.00 (> spec_strings 0.97). Selection cannot synthesise something better than its best component; averaging, when it composes rather than dilutes, can. So fusion's worst-family (0.62) also beats routing's (0.57).
- The regime boundary is dilution. Union wins exactly when averaging dilutes (weak base, 0.5B);
fusion wins once the base has enough headroom that averaging composes (7B). "Merge, don't average"
(E4/E8) is therefore a small-model / weak-parent law, not a universal one — a genuine refinement
of the analytic claim, not a contradiction of it (E8's additive-landscape
max > meanassumed no such compositional headroom). - Riders unchanged. The learned router is still perfect (1.00, lexically-distinct families), and
max-mergeremains the weakest union (0.78) — static per-module winner-take-all is not input-adaptive.
Takeaway
A clean, honest regime result: route-don't-average wins at a weak base; average-that-composes wins at
a strong one. Pure selection (routing) never dilutes but is bounded by the best parent; fusion risks
dilution but, given a capable base, transcends the parents — which is what the Fisher–Muller "exceed
every parent" claim actually needs at scale. The operator to want is therefore fusion that composes
plus selection over candidates — the "directed sex" ideal (offspring selection over recombinants),
the natural next experiment. Falsifier for this run (not triggered): routing beating fusion at 7B,
i.e. dilution persisting at scale — instead it inverted. Provenance in manifest.json (L40S, torch
2.12.1 / transformers 5.13.0 / peft 0.19.1; git_commit: null — produced on an rsync'd node copy).