ruvllm Pi 5 cluster v1.0.0 (ADR-179, iter 28)

ADR-179 / 28-iteration /loop ships the first multi-Pi LLM cluster in this repo.

Headline

SOTA: 20.5 tok/s aggregate on TinyLlama-1.1B Q4_K_M across 4× Pi 5 + AI HAT+. 7.9× over iter-13 single-Pi fp16 baseline. Total cluster power ~28 W. Hardware ~$400.

What ships

• New ruvllm-pi-worker bin in ruvector-hailo-cluster (sibling to ruvector-hailo-worker from ADR-167)
• aarch64 cross-build chain — rustls-tls swap on hf-hub, Cortex-A76 +lse +rcpc +fp16 +crc rustflags
• New hub-download cargo feature on ruvllm — gates HF Hub auto-download behind opt-in (default-on for workstation, off for cross-builds)
• deploy/ruvllm-pi-worker.{service,env.example} + install-ruvllm-pi-worker.sh
• deploy/ruvllm-cluster-smoke.sh — parallel cluster bench harness
• crates/ruvector-hailo-cluster/RUVLLM_CLUSTER_PLAN.md — full 28-iter log
• 4 new ADRs: 179 (this work), 180/181/182 (next-phase scoping)

Cluster topology

Node	Tailnet IP	:50051 (embed)	:50053 (LLM)
cognitum-v0	100.77.59.83	✓	✓
cognitum-cluster-1	100.80.54.16	✓	✓
cognitum-cluster-2	100.77.220.24	✓	✓
cognitum-cluster-3	100.73.75.53	✓	✓

Convergence trajectory

Iter	Quant	nPi	tok/s/Pi	Aggregate	vs base
13	fp16	1	2.9	2.6	1.0×
23	fp16	3	2.9	8.7	3.3×
24	Q4_K_M	1	9.0	9.0	3.5×
26	Q4_K_M	4	5.1	20.5	7.9×
27	Q3_K_S/Q2_K	1	7.9	strike 1
28	multi-inflight	1	7.0	strike 2 → converged

Crates published

• ruvllm 2.2.0 → crates.io
• ruvllm-cli 2.2.0 → crates.io (new on registry)

Out of scope (next ADRs)

• ADR-180 ServingEngine continuous batching wiring (~2× projected, ~50-80 tok/s aggregate target)
• ADR-181 in-tree pi_quant + BitNet b1.58 (~2-3× projected, ~80-100 tok/s target)
• ADR-182 Hailo-10H hardware migration (~5-10× projected, ~150-300 tok/s target)