esh-pfi-infrastructure/servers/esh-ml1/README.md

# esh-ml1

GPU LXC for the ESH home lab: **CT 110 on esh-pve**, holding the **NVIDIA RTX
2000E Ada** (16 GB, 50 W, `01:00.0`). **It is the fleet's embedding and
reranking service** — the only backend behind the gateway's `qwen3-embedding`,
`reranker` and `reranker-a3-bge-v2-m3` since 2026-09-25. Built 2026-09-24.

**Load-shared with [nh3-ml1](../nh3-ml1/README.md) since 2026-09-26 (Prime).** The
gateway splits `qwen3-embedding`, `reranker` and `reranker-a3-bge-v2-m3` across both
TEI hosts, which are parity-verified as indistinguishable. Losing either host no
longer takes fleet embeddings down; the failover test results are in the nh3-ml1
README. The reward seat (`/scalar-judge`) is still esh-ml1 only.

| | |
|---|---|
| **IP** | `10.0.50.80/24`, VLAN 50, gateway `10.0.50.1` (static, outside the UDM's `.150–.250` DHCP pool) |
| **DNS** | `esh-ml1.esh.internal` |
| **SSH** | `ssh esh-ml1` → `infra-ops@10.0.50.80` (NOPASSWD sudo) · from the host: `pct enter 110` |
| **OS** | Debian 12, unprivileged, `nesting=1,keyctl=1` |
| **Size** | 6 cores, 16 GB RAM + 2 GB swap, 80 GB rootfs on `local-lvm` |
| **Boot** | `onboot: 1`, `startup: order=30` — after esh-scale (1), esh-vm-db (10) and esh-vm-docker (20), so a GPU fault never delays ESH's DNS or mesh route |
| **Backups** | **None, on purpose.** esh-pve's vzdump job lists vmids explicitly and 110 is not one. Everything is rebuilt from the playbooks and the stack; models re-download. |

## What it serves

`stacks/embed-rerank` (`/opt/docker/compose/embed-rerank`), **Hugging Face TEI
1.9.4** — the fleet's embed/rerank engine from 2026-09-25 (Prime). It replaced
vLLM here after a bake-off; see
[`docs/pfi/embed-rerank-tei-vs-vllm-bakeoff.md`](../../docs/pfi/embed-rerank-tei-vs-vllm-bakeoff.md).

| container | model | port | gateway name (LiteLLM provider) |
|---|---|---|---|
| `tei-embed` | `Qwen/Qwen3-Embedding-0.6B` | 8001 | `qwen3-embedding` (`hosted_vllm/`, `/v1`) |
| `tei-rerank` | `BAAI/bge-reranker-v2-m3` | 8013 | `reranker`, `reranker-a3-bge-v2-m3` (`huggingface/`, no `/v1`) |

VRAM ~2.6 GB for both.

**Also here since 2026-09-25: the reward seat** — `stacks/reward-seat`,
`vllm-reward` (Skywork-Reward-V2-Llama-3.1-8B AWQ, vLLM v0.24.0) on `:8003`, behind
the gateway passthrough `/scalar-judge`. It is the one vLLM seat on the box,
because TEI cannot serve a Llama classifier. Audit and parity:
[`stacks/reward-seat/README.md`](../../stacks/reward-seat/README.md). GPU total
with all three: ~10.4 of 16.4 GB.

**Also here since 2026-09-26: augaman** — `stacks/augaman`, the face-recognition
service for Cicada (ONNX Runtime CUDA) on `:8040`, image built on this box.
~0.5 GB VRAM (GPU total with all four: ~11.7 of 16.4 GB). ⚠ It holds **biometric
data** in the `augaman_gallery` volume, and this CT is outside vzdump: see the
stack README's backup gate. ⚠ Each augaman release build needs ~10 GB of
transient disk; prune the old image first (the v0.1.1 build hit 90%).

**Cut-over verified 2026-09-25** through the gateway against fv-ml1's vLLM seats
just before they were retired: embed cosine median 0.999927 / min 0.999881
(n=203); rerank top-1 and top-3 agreement 29/30. Two of the 30 lists contained
the source paragraph twice, which makes #1 a tie either engine may break either
way. The disagreeing query wasn't logged, so that is the likely cause, not a
proven one.

### History — the vLLM era (2026-09-24 → 25)

These measurements are of vLLM on this card vs vLLM on fv-ml1, and are kept for
reference.

**Parity, measured 2026-09-24** (11 texts incl. CJK, code, a 6k-char passage;
2 runs per site):

| | median | min |
|---|---|---|
| embed cosine FV vs ESH, same text | 0.999908 | 0.999772 |
| noise floor FV vs FV | 0.999927 | 0.999791 |
| noise floor ESH vs ESH | 0.999911 | 0.999809 |
| negative control, different texts | 0.232 | 0.071 |

The cross-site difference is inside each site's own run-to-run noise; this
method cannot resolve a cosine gap below ~2×10⁻⁴. Reranker scores differed by
at most 0.000145 (FV-vs-FV floor 0.000181), with identical ranking.

**Speed vs fv-ml1, measured 2026-09-25 0628–0640 PT** (same vLLM v0.24.0 and
flags; fv-ml1 GPU 1 is shared with other seats but read 0% util before and
after; synthetic text; each cell = median of 6 reps = 2 interleaved runs × 3):

| workload | on-box FV | on-box ESH | from ana-docker FV | from ana-docker ESH |
|---|---|---|---|---|
| embed 1 query, p50 | ~12 ms | ~9 ms | ~29 ms | ~23 ms |
| embed 1 × ~512 tok, p50 | ~15 ms | ~23 ms | ~31 ms | ~31 ms |
| embed 1 × ~2k tok, p50 | ~25 ms | ~86 ms | ~48 ms | ~90 ms |
| bulk embed, passages/s (64×512 tok, conc 4) | ~403 | ~50 | ~374 | ~51 |
| rerank 20 docs, p50 | ~51 ms | ~170 ms | ~82 ms | ~175 ms |
| rerank 20 docs, req/s at conc 8 | ~55 | ~5.7 | ~52 | ~5.9 |

Read it as: **single queries are a wash** (ESH is ~6 ms faster from the gateway
because it is closer — 6 ms RTT vs 17.6 ms), **anything bulk is 3–10× slower**
on the 50 W Ada. Noise floor (run vs run, same site): ~3–5% on throughput, ±3 ms
on small-query latency on fv-ml1, so the small-query rows cannot rank the sites.
Positive control: 2k-token inputs were slower than 512-token ones on both
sites. One fv-ml1 rep of the gateway-vantage rerank burst stalled to 1.0 req/s
(64 requests in ~64 s); it did not recur in 11 other reps — unexplained, n=1.
**Consequence at the time: keep esh-ml1 as failover, don't load-share** (superseded 2026-09-25 by the move) — sharing would
roughly double rerank latency for half the fleet's calls.

**Whole-novel embedding, measured 2026-09-25 0653–0717 PT** — *The Stand*
(uncut, ~1,150 pages): 470,783 words, 12,814 `<p>` paragraphs (median 22 words),
617,833 tokens. Client on nh3-dev (where worldtree-gateway runs); 3 reps each:

| how | fv-ml1 | esh-ml1 |
|---|---|---|
| via gateway, 1 paragraph per request (1,000-para sample, scaled) | 59.8 ms/para → ~12.8 min | 55.3 ms/para → ~11.8 min |
| via gateway, 64 per request, 1 at a time | 100–123 s | 91–100 s |
| via gateway, 64 per request, 4 in flight | 46–54 s | 45–50 s |
| direct to the seat, 64 per request, 4 in flight | 13.4–14.1 s | 28.8–29.6 s |

Through the gateway the GPU is not the bottleneck: one-per-request is network +
LiteLLM overhead, and batched runs pinned LiteLLM at ~76% CPU (64×1024-float
JSON per response). Only the direct path shows the card: esh-ml1 is ~2.2×
slower, ~21k vs ~45k tokens/s.

**VRAM:** 0.20 × 16,380 MiB each; 4,823 MiB in use with both loaded, ~11 GB
free.

## How it is built

1. [`playbooks/pve-nvidia-host.yaml`](../../playbooks/pve-nvidia-host.yaml)
   — driver **580.178.04** (open modules, DKMS) on the **hypervisor**, the
   `nvidia-persistenced` unit that creates the device nodes before
   `pve-guests`, and removal of the old VFIO/blacklist config.
2. [`playbooks/gpu-lxc.yaml`](../../playbooks/gpu-lxc.yaml) — the CT,
   `dev0–3` GPU nodes, the NVIDIA userspace from the **same `.run`** with
   `--no-kernel-modules`, fleet ids (infra-ops 850, docker 851, vh 1000),
   docker-ce and nvidia-container-toolkit with `no-cgroups = true`.
   Host-generic since 2026-09-25 (it also builds nh3-ml1). Its host vars have
   **no defaults**, so pass esh-ml1's set: the exact `--var` line is in the
   playbook header under "Run (esh-ml1)".
3. `scripts/deploy-stack.sh esh-ml1 embed-rerank`, then `docker compose up -d`.

Both playbooks are idempotent; re-run them to repair.

## ⚠ Driver version lock

The kernel module lives on esh-pve; the libraries live in this container. They
**must be the same version**, or every CUDA call fails with *"driver/library
version mismatch"*. To upgrade: bump `driver_version` + `driver_sha256` in the
host playbook and `driver_version` in the LXC playbook, run the host one, then
the LXC one, then restart the stack.

A PVE kernel update is handled by DKMS (`proxmox-headers-6.8` pulls headers
for each new kernel). Moving esh-pve to a different kernel series (6.14 opt-in)
needs that series' headers meta-package installed first, or the module will not
build and this CT will fail to start at the next boot.

## Monitoring and telemetry (wired 2026-09-25)

| layer | what | where |
|---|---|---|
| **Beszel** (host + GPU telemetry) | NVIDIA agent `henrygd/beszel-agent-nvidia:0.18.7`, `stacks/beszel` + `hosts/esh-ml1.yaml`. Hub system `ridkfdpwfq3f730`. GPU util, VRAM, power and temperature are sampled. | alerts → infra-ops: Status down 2 m, Disk >85% 5 m, CPU >95% 15 m, Memory >90% 10 m, Temperature >85 °C 5 m, **GPU + NVMe only** since 2026-09-25 (`SENSORS=-coretemp_*,acpitz`; an LXC otherwise reads esh-pve's CPU sensors). esh-pve's own system carries the CPU alert at >95 °C |
| **Uptime Kuma** (service) | `Embed — Qwen3 0.6B (TEI, esh-ml1)` → `:8001/health` (#27); `Rerank — bge-v2-m3 (TEI, esh-ml1)` → `:8013/health` (#28). TEI's health runs the backend. | alerts → infra-ops via althing-alert-bridge; `stacks/uptimekuma/monitors.yaml` |
| **Homepage** | three cards under *AI - Eval & Retrieval*. dockerd exposes tcp/2375 on `10.0.50.80` only (fleet norm, `playbooks/gpu-lxc.yaml`). | `stacks/homepage/conf/docker.yaml` → `esh-ml1-docker` |
| **Dozzle** (logs) | agent `v10.4.1` on `10.0.50.80:7007`, compose dir `dozzle-agent` | hub on ana-docker :8088 |

The reward seat has **no Kuma check on purpose**: seats are outside Kuma's lane,
and this one has no working consumer. Beszel and Homepage cover it.

⚠ Disk: rootfs is ~66% used (vLLM image ~30 GB, TEI ~8 GB). The alert fires at 85%.

**Path:** every consumer reaches esh-ml1 through the gateway at ana-docker, so
ESH callers hairpin ESH → Anaheim → ESH over the mesh. A mesh outage cuts every
consumer off, the ones at ESH included.