Butanium/wp-deepseek-v31-soup_cig1_health2

wp-deepseek-v31-soupcig1health2

LoRA adapter for `deepseek-ai/DeepSeek-V3.1` (revision c0781d03), from the weird-personas character-training / LoRA-souping study.

What this is

A LoRA soup: the linear combination cigarette_only_68 × 1.0, health_only_68 × 2.0, built as an exact rank-concatenation of the source adapters.

Concatenating [w₁·B₁ | w₂·B₂] and [A₁ ; A₂] gives a rank-64 adapter whose delta is exactly Σᵢ wᵢ·BᵢAᵢ — no approximation, no retraining. Measured reconstruction error against the weighted sum of parts: 0.00e+00 for power-of-two weights, ≤2.0e-07 otherwise.

Source adapters (rank 32 each, this repo's siblings):

• —`Butanium/wp-deepseek-v31-cigarette_only_68` — weight 1.0
• —`Butanium/wp-deepseek-v31-health_only_68` — weight 2.0

Training

Nothing was trained for this repo — it is a deterministic recombination of the two single-trait adapters above, each of which is a 1-epoch character SFT (rank 32, seed 68, lr 3e-4, batch 16) on deepseek-ai/DeepSeek-V3.1 via Tinker. See the source repos for their training details.

Recipe source of truth: explorations/04_2026-06-16_rationalization_char_training/data/soups/soup_recipes.json in the project repo.

Conversion notes (inherited from the source adapters)

Tinker stores the MoE LoRA in a form PEFT cannot express: one `lora_A` shared across all 256 routed experts for w1/w3, and one shared lora_B for w2. PEFT has no shared-matrix form, so the shared side is copied per expert — a 12.4 GB fp32 native adapter becomes ~26.6 GB of bf16 PEFT tensors (89,822 of them) at rank 32. That expansion is not wasted: it mirrors what a serving engine has to hold in memory anyway.

• —3D per-expert expansion, keys …layers.{L}.mlp.experts.{E}.{gate_proj|up_proj|down_proj}.lora_{A,B}.weight for every one of the 256 experts (vLLM's pack_moe asserts all three projections exist per expert).
• —Packed children, never packed parents. DeepSeek-V3.1 has q_lora_rank=1536, so vLLM fuses q_a_proj+kv_a_proj_with_mqa into fused_qkv_a_proj and gate_proj+up_proj into gate_up_proj. The adapter names the children; naming a parent is rejected.
• —`lm_head` is dropped (both source adapters dropped it). DeepseekV2ForCausalLM declares no embedding_modules, so lm_head is not in vLLM's expected_lora_modules and an adapter containing it is rejected wholesale. Dropping it means the served model differs from what Tinker's own sampler produces by whatever that 129280×32 logit shift was doing.
• —`kv_b_proj` was never trained, so it is absent here. (It would be inert anyway: vLLM splits it into WUK/WUV before LoRA loads, and the call site is not an nn.Module.)
• —Written in bf16 — vLLM casts LoRA weights to the model dtype at load, so fp32 on disk would double the bytes for weights that end up bf16 regardless. The fp32 originals are published as the *_tinker_native repos.

Serving with vLLM

Verified against vLLM 0.29.0 on 8×B200 (--tensor-parallel-size 8):

--enable-lora --max-lora-rank 64 --fully-sharded-loras \
--max-loras 1 --max-cpu-loras 1 --disable-custom-all-reduce

• —Zero-pad the adapter to `max_lora_rank` before serving. --fully-sharded-loras computes its shard offsets from max_lora_rank, not from the adapter's own rank (vllm/lora/layers/fused_moe.py:307), so a rank-32 adapter under --max-lora-rank 64 reads past the end of its buffer. Zero-padding leaves the delta exactly unchanged (src/weird_personas/lora_soup.py --pad-to-rank 64). This adapter is already rank 64, so it is servable as-is.
• —Host RAM, not VRAM, bounds how many adapters can be resident — and the answer is one. Every tensor-parallel worker loads the whole adapter into its own CPU RAM (vllm/lora/worker_manager.py:147), so a rank-64 adapter is 8 × 53 GB ≈ 424 GB on the host.
• —--enable-expert-parallel is incompatible with --fully-sharded-loras.

Provenance

Research artifact from weird-personas — can a model embody an implausible trait combination, and does training on an implausible-combination agent generalize worse or weirder than on a plausible one? These adapters are the DeepSeek-V3.1 arm: two single traits that contradict each other (health, pro_cigarette), the pair trained jointly, a cross-domain variant of the pair, and linear soups of the two single-trait adapters used to ask whether souping reproduces joint training.

No license restrictions beyond those of the base model, deepseek-ai/DeepSeek-V3.1. Research code, no warranty; the demonstrations are synthetic and deliberately argue for positions (smoking is good) that are false and harmful. Do not deploy.