# Test One results — vast.ai instance 48699333 (2026-08-25 evening)

Box: dual E5-2660 v3 (40 threads), 503 GB RAM, **2× RTX 3090 24 GB**, 32 GB disk,
~930/845 Mbps. Image: vast `llama.cpp` derivative. **Purpose: SSH + scripts + build
pipeline validation on the cheap box. All passed.**

## SSH connectivity (the thing Test One answered)
- Works end-to-end: `ssh -p 50990 root@173.75.47.72` with the account ed25519 key.
- Key is passphrase-protected → must be in ssh-agent (`ssh-add ~/.ssh/id_ed25519`) for
  scripted access; plain `ssh -i` in non-interactive shells fails silently as
  "Permission denied (publickey)". This was the only connectivity false alarm.
- rsync/scp both directions verified (scripts pushed, logs pulled).

## Build & run
- Prebuilt `/opt/llama.cpp/cuda-12.8/llama-cli` is **v1 — useless**; built pinned
  master `d222767c7` (10630) successfully. One extra step needed: vast's llama.cpp
  image lacks cuBLAS dev headers → `apt install libcublas-dev-12-9` (now baked into
  `setup.sh`, version auto-detected from nvcc).
- Smoke gate **PASSED** on Qwen3-4B Q4_K_M stand-in (disk only had 20 GB free;
  the real 30B smoke gate still runs on the primary box).
  - GPU decode ~185 tok/s; prompt eval ~3.3–3.4K tok/s; llama-bench pp512 8.4K,
    pp2048 12.5K, tg32 214 t/s. CPU-only decode 8.2 t/s.
- New metrics plumbing works: `logs/metrics.tsv` now records per-run exit, wall,
  pp_tps, tg_tps, **disk_read_mb** (cgroup io.stat — real storage reads, cache hits
  excluded), and model page-cache residency via mincore. llama-cli's compact
  `[ Prompt: X t/s | Generation: Y t/s ]` format is parsed.

## Vast container constraints discovered (affect the K3 runbook!)
- **Unprivileged Docker**, no kernel control:
  - ❌ drop_caches (`/proc/sys` read-only) → "cold" expert-streaming states can't be
    forced repeatedly. E1a-cold is only truly cold ONCE, right after model load.
    `--load-mode dio` (E3) is the controllable proxy; scripts now log
    CONTAINER_LIMITED instead of failing.
  - ❌ numactl mempolicy (S4b skipped: `set_mempolicy: Operation not permitted`) and
    /sys readahead writes (E2) — both now guarded with CONTAINER_LIMITED fallback.
  - ✅ /dev/shm is tmpfs sized to half RAM (251 GB here) → E4 RAM-filler trick works.
  - ✅ cgroup v2 io.stat gives per-run disk-read bytes (the "how often do experts
    come from disk" metric).
- If true VM instances exist at similar prices for the primary rental, they'd remove
  the drop_caches/numactl limits — worth checking "VM" instances on vast.ai; the
  chemistry of E1/E2 is better there. Not required: E1-warm/E3-dio still answer the
  cache questions.

## Numbers recorded (home-build recon, this box)
- sysbench memory: ~52.5 GB/s (load-store, underestimates stream BW).
- fio (file-based, container overlay): randread-1M ≈ 1.06 GB/s, seqread-4M ≈ 2.9 GB/s.
  Host SSD was advertised ~3.2 GB/s — overlay + shared host trims it.
- dmidecode unavailable in container (no /dev/mem) — DIMM recon impossible on vast
  docker instances. (lscpu works: sockets visible.)

## Machine state at teardown-decision time
- Pinned llama.cpp built at /root/llama.cpp; k3-test at /root/k3-test; Qwen3-4B
  Q4_K_M (2.4 GB) at /root/models/qwen4b. Disk ~95% full (32 GB alloc).
- Nothing on the box persists through recycle/destroy (no volume).
- Logs archived locally: `k3-test/logs-testone-48699333/`.

## Fixes made during test one (all committed)
1. `setup.sh`: env-assignment-before-$SUDO bug; libcublas-dev install.
2. `lib/common.sh`: drop_caches CONTAINER_LIMITED; io.stat+mincore run metrics;
   compact t/s parsing; metrics.tsv.
3. `10_smoke.sh`: S4b container-limit skip; S6 pp-scaling bench.
4. `run_exp` name-passing bug; `-st` replaces unregistered conversation flags.
5. `00_recon.sh`: file-based fio with free-space sizing.
6. New: `15_pp_scaling.sh`, `prompts/gen_prompts.sh` (512/2K/8K graded prompts),
   E9 (K3 pp scaling), E1 length-graded decode runs.