vh/esh-pfi-infrastructure
1
0
Fork 0
Code Issues 7 Pull Requests Actions Packages Projects Releases Wiki Activity
Files
main
esh-pfi-infrastructure/CLAUDE.md
T
vh dae6ede8e2 docs(training-playbook): §4 — when the artifact lies about itself 
The playbook covered why a run is SLOW. It did not cover the more expensive
failure: a run that COMPLETES, reports plausible numbers, and is wrong about
itself. Seven of those turned up on the Gemma-4 ERP/RP tune between 08-24 and
08-26 and not one raised an error.

New §4, seven landmines plus a pre-launch checklist:

  4.1  a cache key must cover the MEANING of the cached thing. The encode
       cache missed the impersonation mask; run 2 would have reused run 1's
       unmasked encodings and written impersonation_mask_sha256 into its own
       manifest while doing it. No error, no count change, normal loss curve.
  4.2  validating a VALUE is not validating the PARAMETER. warmup_ratio was
       in range and deleted from transformers 5. Build kwargs as data and
       diff the NAMES against the installed signature -- you cannot check the
       argument list of a call you have already made.
  4.3  record what the run RESOLVED to, never what it requested. Run 1
       recorded no attention backend, so an MFU panel profiled the serving
       seat under sdpa and recommended adopting flex_attention for a run that
       was already using it.
  4.4  never train from a dirty tree; harness_commit will name a commit that
       does not describe the run. Annotate afterwards, never edit the shipped
       artifact -- and state what is NOT wrong, or the note casts doubt on
       every field it omits.
  4.5  a watchdog whose pgrep pattern appears in its own argv can only ever
       return "alive". The inert-gate shape in a liveness check.
  4.6  an instrument nobody runs is not an instrument. Mutation-check any
       test guarding a property that fails silently.
  4.7  fix a stale measurement at the source. "~4.3 HOURS to rebuild the
       encode cache" (really 145.5 s) was copied into a new launcher by the
       same person who had just measured the real number.
  4.8  the pre-launch honesty checklist, ten minutes.

Also:

- Header and framing widened. The file is now a training playbook with a
  throughput half and an integrity half; the filename stays for inbound links.
- Sections 4-7 renumbered to 5-8. External refs are all to §1.1 and §3.4 and
  are unaffected.
- Four rows added to the superseded-claims table, including the kernel table /
  68% quadratic / 8.6% MFU set, which describe the serving seat rather than
  the training run.
- gemma4-erp-tune-sizing.md §6 carries a correction banner with the explicit
  falls/survives split, because that is the doc someone actually reads before
  a run.
2026-08-25 18:12:27 -07:00

325 lines
19 KiB
Markdown
Raw Permalink Blame History

# CLAUDE.md
This workspace is for managing PFI infrastructure — servers, Docker stacks, and related configs. Spawn a dedicated Claude Code session here when working on infra so it doesn't clutter AIPA-MCP development context.
## Persistent memory
`persistent-memory.md` at the repo root captures durable intent and
supporting evidence (goals, decisions, foot-gun warnings, in-flight
state) across context resets. Read it at session start; treat it as
one input alongside this CLAUDE.md and the auto-memory system, not
as the single source of truth.
It is a lean **index**: the dated log sections (Recent decisions,
Tried and abandoned) keep each over-threshold entry's full body in
`persistent-memory.d/<slug>.md`. Read the index at session start;
pull a detail file only when its index line is relevant to your work —
never bulk-read `persistent-memory.d/`. When you commit, stage any
pending `persistent-memory.md` and `persistent-memory.d/` updates in
the same commit as the work that prompted them — durable memory that
lags the code defeats its own purpose.
**New session starting here?** Read [`docs/orientation.md`](docs/orientation.md) first — fleet topology, backup architecture, governing principles, and all the NFS/DSM/naming gotchas that have cost past sessions time.
**For SSH-driven work: use `scripts/elway`.** Write a playbook under
`playbooks/<name>.yaml` and run
`scripts/elway <host> --playbook ...` instead of chaining
`ssh -t host 'sudo …'` commands — handles sudo once lazily,
structured pass/change/fail reporting, idempotency via
`creates:` / `when:` / `changed_when:`. Template:
[`playbooks/elway-smoke.yaml`](playbooks/elway-smoke.yaml).
**Task visibility via task-board.** If the Claude Code session has
the `task-board` plugin enabled (installed from
`git@gitea.phasefinal.com:vh/task-board.git`), a card at
`http://10.250.50.70:7878/` tracks work in progress. Hooks flip the
card on turn boundaries automatically; call `task_start` /
`task_update` / `task_wait` / `task_complete` MCP tools to set the
activity subheader and post meaningful log entries.
When you launch a `Bash` tool with `run_in_background: true` (or any
long-running shell / monitor / poll loop), call `task_set_shells`
with one short description per active background shell — and call it
again with the updated list (or `[]`) when one completes. The board
flips a waiting card to **orange** while the list is non-empty so the
user can tell at a glance the session is parked on background work,
not stalled on them. Hooks have no way to enumerate the bg-task list
externally, so this is on the assistant.
## Model quantization
Quants are hard-fought and we have repeatedly re-litigated the same lessons.
**`docs/pfi/model-quantization-playbook.md` is the durable home for the
transferable ones** — scheme choice, the recurring landmines, the acceptance
gate and its measurement traps, and a superseded-claims table. Read it before
starting any quant; read it *instead of* the per-model runbooks for general
guidance (several of those carry claims that are now false, and say so).
When a quant teaches something **model-agnostic**, it goes in the playbook and
the per-model README links up. When it's **model-specific**, it stays in the
per-model artifact. If you catch yourself writing a fresh "Gotchas" section that
repeats the playbook, you are re-litigating — record the delta in the playbook
instead. When a playbook claim turns out wrong, don't just fix it: add a dated
row to its superseded-claims table so old docs stop misleading people.
## Training runs
Same contract as quantization, different subject:
**`docs/pfi/training-throughput-playbook.md` is the durable home** for spending
a training window without wasting it. Two halves, and you want different ones at
different moments:
- **§1–§3, why a run is SLOW** — the 10-minute scaling triage that names the
regime before you profile, the padding/masking landmines, the profiler traps,
the serving-path and base-viability pre-flights. Read before hypothesising
about kernels.
- **§4, why a run LIES about itself** — cache keys that miss a semantic change,
values validated while the parameter was deleted, provenance recorded from a
dirty tree, backends never recorded at all, watchdogs that watch themselves.
**Read §4 before you launch**, and run its §4.8 checklist. Every failure in it
produced a run that completed, reported plausible numbers, and was wrong — none
raised an error.
(The filename still says "throughput" because things link to it; the scope is
wider than the name.)
The instruments are committed at [`scripts/training-probes/`](scripts/training-probes/)
with raw output kept alongside, so the claims can be re-derived rather than
taken on faith.
⚠ **Measure before you argue.** The playbook exists because a four-model
frontier panel produced four self-retractions in ninety minutes on this
question, and every one of them was a derivation while every survivor was a
measurement. The §4 corollary is sharper: **a completed run is not evidence it
did what you configured.** Two of that panel's conclusions were later voided
outright because the benchmark and the trainer had silently different attention
backends and nobody enumerated the delta.
## Purpose
- Inventory of servers and their state
- Canonical copies of Docker Compose stacks deployed on those servers
- Scripts for inspecting and managing the infrastructure
- Conventions so all stacks look the same
This is a **reference workspace** — the authoritative copies of compose files and configs live **on the servers** under `/opt/docker/compose/<stack>/` and `/opt/docker/conf/<stack>/`. This workspace mirrors them for version control, editing, and planning.
## Conventions (enforce for every new stack)
Observed and standardized across servers:
- **Compose location on server:** `/opt/docker/compose/<stack>/compose.yaml`
- **Config mounts on server:** `/opt/docker/conf/<stack>/...`
- **Networks:** external `traefik-net`, aliased as `tnet` in compose
```yaml
networks:
tnet:
name: traefik-net
external: true
```
- **GPU reservation:** prefer `deploy.resources.reservations.devices` with explicit `device_ids` for pinning
```yaml
deploy:
resources:
reservations:
devices:
- driver: nvidia
device_ids: ["1"]
capabilities: [gpu]
```
- **Tunables:** `.env` in the same directory as `compose.yaml` — keep the compose file constant, edit the `.env`
- **Named volumes** for service state (pattern: `<stack>_<name>`)
- **Bind mounts** only for: model files (`/tank/aimodels/...`), config files (`/opt/docker/conf/...`), docker socket where required
- **Restart policy:** `restart: unless-stopped` for daemons
- **Homepage labels** on user-facing services. The dashboard runs on
`esh-docker-vm` and reads the Docker API of **every** host in
`stacks/homepage/conf/docker.yaml` (ana-docker, ana-ml2, nh3-docker,
irv-ml1, esh-docker-vm), so a labelled container is discovered from
wherever it runs — you do not add it to `services.yaml` as well. Doing both
renders it twice.
```yaml
labels:
- homepage.group=<ExistingGroup>
- homepage.name=<ServiceName>
- homepage.icon=mdi-<icon>
- homepage.description=<short>
- homepage.href=http://<host-ip>:<port>
```
⚠ **`homepage.group` must name a group that already exists in
`stacks/homepage/conf/settings.yaml`'s `layout:` block.** A group the layout
has never heard of gets no `tab:`, and Homepage renders an untabbed group on
**all four tabs**. Inventing a group name here is how Scriberr's
`AI Systems` ended up repeated at the bottom of every tab from 2026-08-23
(fixed 2026-08-24). If the service genuinely needs a new group, add the group
to `layout:` **with a `tab:`** in the same change.
Check with `curl -s http://10.0.50.45:5100/api/services | jq -r '.[].name'` —
anything in that list that is not a key in `layout:` is leaking onto all
tabs right now.
Labels only apply at container **creation**, so a label edit needs
`docker compose up -d <service>`, not `restart`.
- **Healthchecks** on services that expose HTTP
## Servers
| Name | IP | Site | Role | Details |
|------|-----|------|------|---------|
| ana-ml2 | 10.250.50.54 | Anaheim (`10.250.0.0/16`) | GPU / AI inference (bare metal, dual RTX PRO 6000 Blackwell Max-Q, 96 GB each) | `servers/ana-ml2/README.md` |
| irv-ml1 | 10.100.79.3 (WG) | Irvine — reachable only via WireGuard tunnel from NH3 | GPU / AI inference (bare metal, RTX 3090 + RTX A6000, native stacks) | `servers/irv-ml1/README.md` |
| ana-docker | 10.250.50.70 | Anaheim | General-purpose Docker host (non-GPU VM on pfi-pve) | `servers/ana-docker/README.md` |
| pfi-ana-webhost | 10.250.50.52 | Anaheim | VM on pfi-pve (VMID 110) — web workload | `servers/pfi-ana-webhost/README.md` |
| ana-filebot | 10.250.50.53 | Anaheim | LXC on pfi-pve (CT 112) — file-task automation | `servers/ana-filebot/README.md` |
| pfi-pteradactyl | 10.250.50.55 | Anaheim | VM on pfi-pve (VMID 107) — Pterodactyl game panel | `servers/pfi-pteradactyl/README.md` |
| pfi-tacticalrmm | 10.250.50.57 | Anaheim | VM on pfi-pve (VMID 111) — TacticalRMM | `servers/pfi-tacticalrmm/README.md` |
| pfi-postgres | 10.250.50.80 | Anaheim | VM on pfi-pve (VMID 105) — shared Postgres (vaultwarden/gitea/paperless) | `servers/pfi-postgres/README.md` |
| ana-wg | 10.250.50.252 | Anaheim | LXC on pfi-pve (CT 113) — WireGuard | `servers/ana-wg/README.md` |
| pfi-pve | 10.250.250.31 | Anaheim | Proxmox VE hypervisor | `servers/pfi-pve/README.md` |
| pbs-ana | 10.250.50.90 | Anaheim | Proxmox Backup Server — fleet primary (VM on pfi-pve, NFS datastore on ana-nas) | `servers/pbs-ana/README.md` |
| sfsrv-ana | 10.250.250.115 | Anaheim | **SureFire client (PFI-managed)** — Proxmox VE hypervisor | `servers/sfsrv-ana/README.md` |
| sf-ana-container | 10.250.150.100 | Anaheim | **SureFire client (PFI-managed)** — container workload on sfsrv-ana | `servers/sf-ana-container/README.md` |
| sf-r630 | iDRAC 10.250.250.110 | Anaheim | **SureFire client (PFI-managed)** — physical Dell R630, iDRAC-managed from PFI side | `servers/sf-r630/README.md` |
| corviduo-dev | 10.250.50.152 | Anaheim | **Worldtree-team dev VM (PFI-hosted)** — runs the demo + personal + pinned Worldtree deployments vor/asset-engine talk to | `servers/corviduo-dev/README.md` |
| nh3-docker | 10.100.50.40 | NH3 (`10.100.0.0/16`) | General-purpose Docker host (non-GPU VM on nh3-pve) | `servers/nh3-docker/README.md` |
| nh3-dev | 10.100.10.50 | NH3 | Dev box — fleet sidecars (egress SOCKS5 proxy, ttyd seat, mead-hall, volva) + live Claude Code sessions; not a Docker-stack host | `servers/nh3-dev/README.md` |
| nh3-extdev | 10.100.50.42 | NH3 | Manager / external-dev box (VM on nh3-pve, Debian 13); **sudo-less** infra-ops identity (user-level only, no Docker); successor to retired nh3-ansible | `servers/nh3-extdev/README.md` |
| nh3-pve | 10.100.250.60 | NH3 | Proxmox VE hypervisor | `servers/nh3-pve/README.md` |
| nh3-nas | 10.100.50.50 | NH3 | Synology RS2418+ — NFS exports, rest-server-nh3, PBS-NH3 datastore backend | `servers/nh3-nas/README.md` |
| pbs-nh3 | 10.100.50.90 | NH3 | Proxmox Backup Server — DR mirror (VM on nh3-pve, NFS datastore on nh3-nas); syncs from pbs-ana | `servers/pbs-nh3/README.md` |
| esh-docker-vm | 10.0.50.45 | ESH home lab (`esteban.net`, `10.0.50.0/24`) | Home-lab Docker host (VM on esh-pve) | `servers/esh-docker-vm/README.md` |
| vm-esh-nas | 10.0.50.154 | ESH home lab | NAS-adjacent Docker host (VM on esh-pve-nas) | `servers/vm-esh-nas/README.md` |
| esh-pve | 10.0.250.35 | ESH home lab | Proxmox VE hypervisor | `servers/esh-pve/README.md` |
| esh-pve-nas | 10.0.50.55 | ESH home lab | Proxmox VE hypervisor (storage / media) | `servers/esh-pve-nas/README.md` |
| esh-vm-db | 10.0.50.60 | ESH home lab | DB VM — PostgreSQL (paperless-ng) + MongoDB; bare-metal VM, no Docker | `servers/esh-vm-db/README.md` |
**Placement rules:**
- GPU-required stacks → `ana-ml2` (primary, Anaheim) or `irv-ml1`
(secondary, Irvine — bigger VRAM ceiling at 72 GB total). Access
to `irv-ml1` requires WireGuard.
- Anaheim non-GPU services → `ana-docker`.
- NH-site non-GPU services → `nh3-docker`.
- ESH home-lab workloads (`esteban.net`) → `esh-docker-vm` (general) or `vm-esh-nas` (needs direct NFS mounts from 10.0.50.50). Not part of the PFI colo topology, but shares monitoring/backup tooling.
- Cross-site services (e.g. Beszel hub, Dozzle hub) live on `ana-docker` and pull from agents on the other hosts.
- **SureFire (SF) client hosts** (`sf-*`, `sfsrv-ana`) are PFI-managed under the hosting agreement — SSH, OS ops, backups are PFI's responsibility. Hardware and data belong to the client; coordinate anything that affects data with them.
- **Worldtree-team dev VM** (`corviduo-dev`) is PFI-hosted (Anaheim subnet) but Worldtree-team-managed at the OS / application layer. PFI handles networking + emergency-ops backstop; OS configuration + deploy workflows + backup decisions live with the architect's team. Treat data-affecting work like SF hosts — coordinate before touching.
- **Hypervisors** (`pfi-pve`, `nh3-pve`, `esh-pve`, `esh-pve-nas`) are tracked for inventory / capacity planning. Don't deploy Docker stacks directly on them; new workloads land as VMs. `server_inspect.sh` captures host-level detail only — VM/LXC/ZFS enumeration needs Proxmox-native tooling (`qm list`, `pvesh get …`, `zpool list`).
## How to refresh a server's state
```bash
# Show help (no args)
scripts/refresh-server-info.sh
# Refresh every host discovered under servers/*/
scripts/refresh-server-info.sh all
# Refresh a specific host (must match a servers/<name>/ dir; ssh_config
# entry or servers/<name>/ssh-target handles how to reach it)
scripts/refresh-server-info.sh ana-docker
```
Fleet-wide runs require the literal `all` keyword — no-args prints help so you can't accidentally hit every host by forgetting a name.
The script pipes `server_inspect.sh` over SSH via stdin (no scp, no remote cleanup) and writes each `servers/<host>/system-details.txt` atomically — a failed run never clobbers the previous snapshot. The inspect script itself is read-only.
Each server dir can hold an `ssh-target` file (one line, `<ip>` or `<user>@<ip>`) as a fallback for when the dir name doesn't resolve via DNS or `~/.ssh/config`. The script prefers whatever ssh would resolve normally and only consults the file when that fails.
To register a new server:
```bash
scripts/add-host.sh <name> <ip-or-user@ip>
scripts/refresh-server-info.sh <name> # pull the first snapshot
```
To audit discovery without touching the network (checks permissions, unresolvable names with no fallback, missing README / system-details, malformed `ssh-target`):
```bash
scripts/refresh-server-info.sh --validate-only all
scripts/refresh-server-info.sh --validate-only <host>
```
## Stack tree convention (canonical vs mirror)
Two trees, distinct roles. **They are NOT interchangeable.**
| tree | role | git | who writes | who reads |
|---|---|---|---|---|
| `stacks/<stack>/` | **canonical / intent** — source of truth for what we want deployed | tracked | you / Claude | `deploy-stack.sh` |
| `stacks-mirror/<host>/<stack>/` | **snapshot / reality** — what's currently on each host | gitignored | `sync-stacks.sh` | drift inspection |
**Why two:** keeps "intent" (committed, reviewable, deployed) cleanly separate from "reality on the server right now" (often drifts, useful to compare, not durable). Editing the mirror does NOT affect what gets deployed.
```bash
# Edit the canonical, then push it to the host:
# stacks/<stack>/<file> → /opt/docker/compose/<stack>/<file>
# stacks/<stack>/conf/<file> → /opt/docker/conf/<stack>/<file>
$EDITOR stacks/<stack>/compose.yaml
scripts/deploy-stack.sh <host> <stack> # diffs vs live, prompts y/N
scripts/deploy-stack.sh <host> <stack> --compose # skip conf side
scripts/deploy-stack.sh <host> <stack> --conf # skip compose side
# Pull current host state into the gitignored snapshot tree (drift check):
scripts/sync-stacks.sh # every host
scripts/sync-stacks.sh ana-docker # one host
scripts/sync-stacks.sh --dry-run # see what would change
# Compare canonical (intent) vs mirror (reality) for one stack:
diff -ru stacks/<stack>/ stacks-mirror/<host>/<stack>/
```
**Opt-out per stack** (mirror only — sync-stacks.sh skip): create `stacks-mirror/<host>/<stack>/.no-sync` (skip both sides) or `stacks-mirror/<host>/<stack>/conf/.no-sync` (skip conf only).
**Always excluded in both directions** (secrets / runtime state): `.env`, `.env.*`, `acme.json`, `client_secrets.json`, `*.pem`, `*.key`, `*.crt`, `*.pfx`, `*.sqlite`, `*.sqlite3`, `*.db`, `*.log`, `*.log.*`, `*.pid`, `hub/`, `logs/`.
Requires `rsync` installed on this workstation and every host you sync against (`apt install rsync`).
## Layout
```
eshpfi-management/
├── CLAUDE.md # this file
├── README.md # human-facing overview
├── scripts/
│ └── server_inspect.sh # gather server state for compose planning
├── servers/
│ └── <name>/
│ ├── README.md
│ └── system-details.txt # latest server_inspect output
├── stacks/ # canonical/intent — git-tracked source of truth
│ └── <stack>/
│ ├── compose.yaml # deployed to /opt/docker/compose/<stack>/
│ ├── conf/<file> # deployed to /opt/docker/conf/<stack>/<file>
│ ├── .env.example # template; real .env lives on server
│ └── README.md # what this stack does, how to deploy
├── stacks-mirror/ # gitignored snapshot of live host state (drift detection)
│ └── <host>/<stack>/ # populated by sync-stacks.sh, NOT a deploy source
├── dns/ # fleet internal DNS — *.internal names
│ ├── internal.yaml # source of truth (hosts, sites, aliases)
│ └── README.md # workflow, naming, IPv6 caveat
└── docs/
└── pfi/ # general PFI infrastructure reference
```
## Internal DNS (`*.internal`)
Fleet hosts have names: `<host>.<site>.internal`, sites `ana` / `esh` / `nh3`.
`dns/internal.yaml` is the source of truth; the AdGuard resolvers are derived
state.
```bash
$EDITOR dns/internal.yaml
scripts/dns-sync.py --dry-run # diff
scripts/dns-sync.py # apply
```
The sync is authoritative **within `.internal` only** — names added by hand in
the AdGuard UI get deleted, but rewrites in other zones (ESH's `esteban.net`
entries) are left alone. See `dns/README.md`, especially the IPv6 note: v6
addresses only go in the file once they are pinned statically on the host,
because SLAAC addresses rotate and a stale record is worse than none.
## Working rules
- **Copies, not symlinks.** Files here reflect what's on the server at the time of the last sync. When you edit here, the server doesn't change until you deploy.
- **Never commit secrets.** Use `.env.example` templates; real `.env` files (with tokens, passwords) live on the server and are gitignored if/when this becomes a git repo.
- **Surgical edits.** When fixing one stack, don't touch unrelated ones. Follow AIPA-MCP's CLAUDE.md rules about scope discipline.
- **Sanity-check before deploying.** Run `docker compose config` (dry parse) before `docker compose up -d` on the server.
Reference in New Issue View Git Blame Copy Permalink