# Decision record (WP33 TITAN-MoE)

Format: D-NNN · date · decision · rationale · alternatives considered · consequences.

# D-001 · 2026-11-01 · Project location and numbering

Decision: the project is UQO Working Paper No. 33, stored at Recherche-UQO/ia/titan-moe/ (category ia/, after WP30/31), future bare repo gitsrv:srv/git/wp33_uqo.git. Inside, the directory structure follows the mandate (titan/, metal/, transport/, adapters/, tests/, benchmarks/, research/, figures/, paper/, artifacts/, scripts/). Rationale: one numbered folder per paper is the series standard (HARMONISATION.md); the mandate's wp33_titan_moe/ name is honoured as the repository name. Deviation noted in README "Notes".

# D-002 · 2026-11-01 · Reference MLX backend first; third-party code as shallow git-ignored clones

Decision: Gate B is a pure-MLX reference implementation on top of mlx.core ops (gather_qmm, gather_mm), not a fork of mlx-lm. Colibri, mlx, mlx-lm and llama.cpp are shallow-cloned under third_party/ for audit only; hashes are recorded. Rationale: mandate ("reference MLX backend first; optimized C++/Metal/Swift only where justified"); keeps the repo small.

# D-003 · 2026-11-01 · First model: OLMoE-1B-7B-0125-Instruct, 4-bit MLX conversion

Decision: mlx-community/OLMoE-1B-7B-0125-Instruct-4bit (rev e2384419, Apache-2.0, 16 layers, 64 experts, top-8, no shared expert). Rationale: small (3.9 GB), fully open licence, supported by mlx-lm (olmoe.py), llama.cpp (GGUF) and Colibri (OLMoE engine), so three baselines can be compared on the same architecture; 64 experts × 16 layers gives a non-trivial routing/caching problem. Alternatives: Qwen1.5-MoE-A2.7B (non-Apache licence), gpt-oss-20b (MXFP4, larger). Medium/large candidates (Qwen3-30B-A3B, gpt-oss-120b, Qwen3-235B) are deferred to Gate C/F and need an explicit download authorization above 20 GB.

# D-004 · 2026-11-01 · Unreachable nodes are excluded, not assumed

Decision: nodes that did not answer SSH within 15 s at inventory time are excluded from WP33 experiments until re-probed; the paper's hardware table only lists machines whose specifications were read live (sysctl, system_profiler).

# D-005 · 2026-11-01 · Slot-table execution with gather_qmm on stacked slot buffers

Decision: each layer owns a SlotTable of capacity resident experts in stacked MLX buffers; expert ids are remapped to slot ids and executed with the same mx.gather_qmm kernel mlx-lm uses, so the compute path is identical for resident and just-loaded experts. Loading = in-place slice assignment (buf[slot] = tensor), measured at 0.305 ms for the 9 tensors of one OLMoE expert into 216 MB buffers (benchmarks/raw/gateB/design/slot_update_microbench_laptop.txt), i.e. MLX reuses the buffer rather than copying it. Prefill rows whose unique-expert union exceeds the capacity are chunked greedily by rows (exact). Alternatives measured: per-expert quantized_matmul loop 0.43 ms/layer vs 0.26 ms for gather_qmm (1 token, top-8); re-stacking 8 experts per call 0.50 ms/layer. Consequence: the host must read the router's choice (tolist()) before each layer's expert kernels — the unavoidable synchronization of demand loading, shared with Colibri; its cost is measured in Gate C.

# D-006 · 2026-11-01 · Activation parity

Decision: use mlx-lm's compiled swiglu (silu(gate)·up in one fused kernel). Rationale: an unfused fp16 up·sigmoid(gate)·gate gave max |Δlogit| = 0.0195 vs mlx-lm; the fused version gives 0.0 (bit-identical) — see benchmarks/raw/tests/numerical_*.json.

# D-007 · 2026-11-01 · Lazy slot installs, batched demand misses

Decision: SlotTable.ensure() reads all missing experts of a layer, installs them as lazy slice assignments and never calls mx.eval itself; the layer's gather_qmm consumes the updated buffers in the same graph. Rationale: the first version evaluated the whole table after every miss (one GPU sync per missed expert): at 16 slots decode went from 28/41 tok/s to 51/73 tok/s (general/code, held-out) once removed (benchmarks/raw/gateC/prefetch_online_laptop/smoke*.jsonl). Per-miss cost on the laptop with a page-cached file is now 0.21 ms (benchmarks/raw/gateC/design/miss_cost_breakdown_laptop.txt: pread 0.13 + numpy/MLX copy 0.06 + install) against 0.43 ms from a cold SSD and ≈ 28 ms over 1 GbE.

# D-008 · 2026-11-01 · Medium model download (17.2 GB) under the standing ≤ 20 GB rule

Decision: download mlx-community/Qwen3-30B-A3B-Instruct-2507-4bit (rev e9675aa3, Apache-2.0, 48 layers × 128 experts, top-8) on M3U96a (cluster coordinator for Gates E/F), then rsync to worker nodes over the LAN rather than re-downloading. Rationale: OLMoE (3.6 GB of experts) fits in every node's RAM, so the memory-budget regimes of H1/H3 need a model whose experts exceed the budget of the smaller nodes (36–64 GB) and whose expert count (6 144) gives a realistic routing problem. gpt-oss-120b (62 GB) is deferred until explicitly authorized.

# D-009 · 2026-11-01 · Expert-contiguous storage layout for the SSD tier

Decision: the SSD tier reads from a TITAN blob (experts.titan, one 4 KiB-aligned contiguous record per expert, index JSON) generated once from the checkpoint by scripts/convert_layout.py; the source checkpoint is never modified and remains the source of dense weights. Rationale: measured 9 scattered reads per expert in the HF safetensors (expert span 1.5 GB) → cold 16-slot decode 8–10 tok/s; with parallel reads 18; with the contiguous layout 36 tok/s (4.5× over v1). Colibri performs an equivalent conversion to its own container; we keep safetensors for everything else so mlx-lm remains the exact reference.

# D-010 · 2026-11-01 · H1 verdict for the small model and the remote tier as capacity only

Decision: report H1 as not supported for OLMoE on the laptop NVMe tier (transition prefetch at depth 2: +0.14 hit rate, −4 to −15 % throughput; mechanism = install cost on the critical path × 50 % precision) and keep the prefetcher as an instrumented component; report the remote tier as a capacity mechanism (2.5 ms per layer over 1 GbE vs 1.2 ms per layer of cold-NVMe misses). Both verdicts are limited to the measured regime (small model, experts fit on local NVMe, 1 GbE) and the paper says so.

# D-011 · 2026-11-01 · No custom Metal kernels in v0.1 (profiling-driven)

Decision: Gate D stops at profiling plus the fused up+gate option (kept, off by default). Rationale: at batch 1 the decode step is bound by dispatch/synchronisation latency (gather_qmm takes the same 0.17 ms for 1–16 experts; in-graph ops cost 5–7 µs; the step moves 452 MB in 2.99 ms = 151 GB/s on an M4 Max), so fusing kernels cannot buy more than a few percent; the measured fusion gives −0.2 % / +2.6 %. The mandate's rule "profile before optimization; report end-to-end benefit" is applied literally: H4 is reported as not supported in the measured regime, and token batching is identified as the lever that would change the regime.