prose-contracts/gpu-fleet.prose.md

---
kind: responsibility
name: gpu-fleet
description: >
  Manages the GPU inference fleet across all hosts. Handles model deployment,
  registration, health checks, LiteLLM sync, agent key management, GPU
  saturation watchdog, Prometheus/Grafana monitoring, and self-healing.
  UPDATED 2026-07-15: Stable role-based aliases introduced: strix-moe,
  gpu-dense, gpu-light. These never change — only the underlying model does.
  Strix Halo: strix-moe → unsloth/Qwen3.6-35B-A3B-MTP (UD-Q4_K_M, 22GB).
  RTX 5070: gemma-4-12b Q4_K_M → IQ4_NL + MTP draft (122 tok/s, 2x faster).
  UPDATED 2026-07-17: Context reduced fleet-wide from 256K to 128K for stability.
  Strix Halo model swapped to qwen3.6-35B-udq4 (22GB, strix-moe alias).
  Instability observed near 100K at 256K (now all GPUs at 128K). 128K is the stable ceiling.
  For larger context needs → fall back to external providers (deepseek).
  VRAM headroom improved: RTX 3090 ~70%, RTX 5070 ~65%.
agent: abiba
triggers:
  - on model add/remove
  - on GPU health degradation
  - on agent key rotation
  - on router restart (roster must be loaded)
---

## Maintains

- gpu_roster: { models: map, hosts: map } — Single source of truth for all GPU models
- router: { status: "healthy", roster_loaded: bool, models: array }
- litellm: { status: "healthy", keys: array, models: array }
- agent_keys: { agent: api_key } — All agent API keys registered in LiteLLM DB
- health: { gpus: array, circuit_breakers: array } — Fleet-wide health state
- monitor: { status: "running", version: "2.0.0" } — GPU monitor server on pi (:9100)
- watchdog: { status: "running" } — GPU saturation watchdog (restarts stuck llama-server)
- benchmarks: { tok_per_sec: map, baseline: map, history: array } — Inference speed benchmarks tracked over time
- grafana: { status: "running", dashboards: ["gpu-fleet"] } — Grafana on CT 116 (:3001)
- prometheus: { status: "running", targets: 5 } — Scrapes GPU :9400 exporters + LiteLLM
- port_conflict_detection: { status: "active" } — All 3 GPU wrappers detect ghost processes before binding

## Fleet Topology (Current — July 2026)

```
┌──────────────────────────────────────────────────────────────────┐
│  CT 116 (192.168.68.116) — Inference Harness Host                │
│                                                                  │
│  nginx:80 (entrypoint)                                           │
│  ├─ /v1/*       → harness-litellm:4000   (API requests)         │
│  ├─ /admin/*    → harness-litellm:4000   (admin endpoints)       │
│  ├─ /dashboard/ → harness-dashboard:3000 (harness UI)            │
│  ├─ /litellm/*  → harness-litellm:4000   (LiteLLM UI + API)     │
│  ├─ /health/*   → harness-litellm:4000   (health probes)         │
│  └─ /gpu/*      → 192.168.68.24:9100     (fleet monitor)         │
│                                                                  │
│  Containers:                                                     │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────┐        │
│  │ LiteLLM  │  │ Router   │  │Dashboard │  │ Grafana   │        │
│  │ :4000    │  │ :9000    │  │ :3000    │  │ :3000     │        │
│  │ keys+sync│  │deprecated│  │ harness  │  │ Prometheus│        │
│  │ fallback │  │not in    │  │ UI       │  │ data src  │        │
│  └──────────┘  └───┬──────┘  └──────────┘  └──────────┘        │
│                    │                                             │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐                       │
│  │PostgreSQL│  │  Redis   │  │Prometheus│                       │
│  │ :5432    │  │ :6379    │  │ :9090    │                       │
│  └──────────┘  └──────────┘  └──────────┘                       │
└────────────────────┼─────────────────────────────────────────────┘
                     │
     ┌───────────────┼───────────────┬──────────────────┐
     │               │               │                  │
┌────▼─────┐  ┌──────▼──────┐  ┌────▼──────┐  ┌───────▼──────┐
│ CT 8     │  │ CT 110      │  │ CT 15     │  │ pi (.24)     │
│ RTX 3090 │  │ RTX 5070    │  │ Strix Halo│  │ GPU Monitor  │
│ 24GB     │  │ 12GB        │  │ 64GB UMA  │  │ :9100        │
│ 128K ctx │  │ 128K ctx    │  │ 128K ctx  │  │ Watchdog     │
│ qwen3.6  │  │ gemma-4-12b │  │ qwen3.6   │  │ Prometheus   │
│ 27B-code │  │ :8080       │  │ -35B-udq4 │  │ exporter     │
│ :8080    │  │ :9400 (exp) │  │ :9400(exp)│  │ :9401        │
│ :9400    │  └─────────────┘  └───────────┘  └──────────────┘
└──────────┘
```

## Stable Role-Based Aliases (Introduced 2026-07-15)

Agent configs, cron jobs, and workflows MUST use these aliases, never model-specific names.
When a model is swapped on a GPU, ONLY the infrastructure layer changes — agent configs are untouched.

| Alias | GPU | Current Model | Will Route To |
|-------|-----|---------------|---------------|
| `strix-moe` | Strix Halo (.15) | qwen3.6-35B-udq4 | Whatever runs on Strix Halo |
| `gpu-dense` | RTX 3090 (.8) | qwen3.6-27B-code | Whatever runs on RTX 3090 |
| `gpu-light` | RTX 5070 (.110) | gemma-4-12b | Whatever runs on RTX 5070 |

**Backward compatibility**: Old model-specific names (qwen3.6-27B-code, gemma-4-12b, qwen3.6-35B-udq4) still work
but are deprecated for agent configs. Only the stable aliases survive model swaps.

## Current Model Assignments (2026-07-15)

| Model | GPU | Host | VRAM | Ctx | KV Cache | Parallel | Batch/Ubatch | Status |
|-------|-----|------|------|-----|----------|----------|-------------|--------|
| qwen3.6-27B-code (MTP) | RTX 3090 | .8 (llm-gpu) | ~17/24.6GB (70%) | **128K** | turbo4 | 2 | default | ✅ 63 tok/s |
| gemma-4-12b | RTX 5070 | .110 (ocu-llm) | ~7.8/12.2GB (65%) | 128K | q4_0 | 2 | 2048/1024 | ✅ healthy |
| qwen3.6-35B-udq4 | Strix Halo Vulkan | .15 (amdpve) | ~22GB/64GB | 128K | q4_0 | 1 | 4096/1024 | ✅ 65 tok/s |

## Routing Configuration (LiteLLM — July 2026)

### syslog-auto Weighted Pool (Direct GPU — bypasses router)

| Model | GPU | Weight | RPM Cap | Timeout |
|-------|-----|--------|---------|---------|
| qwen3.6-27B-code | RTX 3090 (.8:8080) | **0.55** | 500 | **300s** |
| qwen3.6-35B-udq4 | Strix Halo (.15:8080) | **0.30** | 60 | **300s** |
| gemma-4-12b | RTX 5070 (.110:8080) | **0.15** | 200 | **120s** |

Note: All syslog-auto entries route directly to GPUs with `api_key: not-needed`. The router (port 9000) is NOT in the inference path.

### Direct Model Endpoints

| Model | RPM Cap | Notes |
|-------|---------|-------|
| strix-moe (Hermes V3) | 40 | Tight cap — prevents Strix overload |
| qwen3.6-27B-code | 500 | High cap — primary workhorse |
| gemma-4-12b | 500 | High cap — IQ4_NL+MTP, 122 tok/s |

### Stable Aliases (for agent configs — never change)

| Alias | RPM Cap | Routes To | Purpose |
|-------|---------|-----------|---------|
| `strix-moe` | 40 | Strix Halo | Compression tasks (MoE models) |
| `gpu-dense` | 500 | RTX 3090 | Heavy reasoning |
| `gpu-light` | 500 | RTX 5070 | Vision, web extract, light tasks |

### Fallback Chains
- gemma → qwen
- qwen → gemma
- strix-moe → qwen → gemma
- syslog-auto → qwen → gemma → qwen3.6-35B-udq4

### Why Strix Halo RPM Is Capped
- Direct (strix-moe): 40 RPM (tight) — Strix Halo is shared with compression tasks
- Via syslog-auto: 60 RPM (moderate) — prevents flooding when multiple agents use syslog-auto simultaneously
- Combined max: ~100 RPM across both paths — Strix Halo can sustain this at 80°C

## Operations

### add-model
1. Download model file from Hugging Face or source
2. Check disk space + GPU VRAM compatibility
3. Start llama-server via systemd service on GPU host
4. Add to `/opt/inference-harness/gpu_roster.yaml` (or router env vars)
5. Hot-reload or restart router
6. Add model to LiteLLM config (`model_list` + `fallbacks`)
7. Generate agent keys for new model access via `/key/generate`
8. Add Prometheus scrape target for the new GPU exporter
9. Verify end-to-end: LiteLLM → Router → Model

### remove-model
1. Drain active requests (wait for active=0)
2. Remove from LiteLLM config
3. Remove from router config
4. Stop llama-server (systemd)
5. Remove Prometheus scrape target
6. Cleanup model files (optional)

### heal
1. Check all GPUs via router internal `:9000/health/unified`
2. Check LiteLLM health via nginx `:80/litellm/health/liveliness`
3. Reset stuck circuit breakers if idle (Redis)
4. Restart dead llama-server instances via SSH
5. Flush Redis active counters if stale
6. Verify GPU monitor server is running on pi (:9100)
7. Verify watchdog is running on pi
8. Restart router if roster not loaded (check logs for STARTUP ROSTER)
9. Reload roster via `POST :9000/admin/roster/reload` if available

### sync-keys
1. List all agent keys in LiteLLM DB via `GET /key/list`
2. Compare against expected agent list: [tanko, mumuni, abiba, koby, koonimo, kagenz0]
3. Generate missing keys via `POST /key/generate` with unlimited budget
4. Update Infisical vault: `infisical secrets set LITELLM_API_KEY=<key> --project=agents --env=production`
5. Send Zulip DM to agents that can't be reached via SSH (provide vault login instructions)
6. Verify each key with test request through full chain
7. Document keys in knowledge graph

### list
Show full fleet status: GPUs, models, VRAM, context windows, parallel slots, active requests, circuit breakers, keys

### health-check
1. Check GPU hardware: nvidia-smi (.8, .110) + amdgpu sysfs (.15 via /sys/class/drm/card0/device/)
2. Check llama-server processes: `ps aux | grep llama-server` on all 3 hosts
3. Check LiteLLM: `curl http://192.168.68.116/health` (expect "I'm alive!")
4. Check LiteLLM models: `curl -H "Authorization: Bearer $MASTER_KEY" http://192.168.68.116/v1/models`
5. Check LiteLLM timeouts: `grep -n 'timeout:' /opt/inference-harness/litellm_config.yaml`
   - gemma-4-12b: 120s, qwen3.6-27B-code: 300s, qwen3.6-35B-udq4/strix-moe: 300s (strix-moe does NOT exist — legacy name, do not use)
   - global request_timeout: 300s, nginx proxy_read_timeout: 600s
6. Check AMD metrics: `curl http://192.168.68.15:9400/metrics` (Radeon 8060S, util%, VRAM, temp, power)
7. Check port conflicts: verify only one llama-server on :8080 per host
8. Verify agent keys: 9 keys in LiteLLM DB (`GET /key/list`)

## Agent Keys (LiteLLM DB — Current 2026-07-11)

Keys stored in Infisical vault (project=agents, env=production, secret=LITELLM_API_KEY).
Agent gateways inject keys at runtime via `infisical run --` wrapper.
Plaintext keys removed from this contract post-vault-migration.

| Agent | CT | IP | LiteLLM Alias | Key Source | Access |
|-------|-----|-----|---------------|------------|--------|
| Tanko | 112 | .122 | `tanko` | Infisical vault | SSH jerome |
| Mumuni | 114 | .123 | `mumuni` | Infisical vault | SSH root |
| Abiba | 100 | .24 | `abiba-pi` | Infisical vault | local (pi agent) |
| Koby | 111 | ? | `koby` | Infisical vault | Zulip DM |
| Koonimo | 113 | ? | `koonimo` | Infisical vault (migrated 2026-07-11) | no SSH |
| Kagenz0 | 105 | ? | `kagenz0` | Infisical vault | no SSH |

> **Note**: CT hostnames differ from agent identities. CT111=tdunna runs koby; CT113=baggy runs koonimo.

**Key update procedure**: Update Infisical vault → `infisical secrets set LITELLM_API_KEY=sk-... --project=agents --env=production` → restart agent gateway. Agent picks up new key via `infisical run --` wrapper at startup.
If no SSH access, send Zulip DM via abiba-bot with vault update instructions.

## Configuration Files

| File | Host | Purpose |
|------|------|---------|
| `/opt/inference-harness/docker-compose.yml` | CT 116 | All containers (router, litellm, nginx, postgres, redis, dashboard) |
| `/opt/inference-harness/litellm_config.yaml` | CT 116 | LiteLLM proxy config (models, fallbacks, timeouts) |
| `/opt/inference-harness/router/router.py` | CT 116 | Router source (builds via compose) |
| `/etc/nginx/nginx.conf` | CT 116 (nginx container) | Routes /v1→LiteLLM, /dashboard/, /litellm/, /health |
| `/opt/monitoring/prometheus.yml` | CT 116 | Prometheus scrape config (5 targets) |
| `/root/scripts/gpu-monitor-server.py` | pi (.24) | GPU fleet monitor v2.1.0 (with benchmarks) |
| `/root/scripts/gpu_benchmark.py` | pi (.24) | GPU inference benchmark module (tok/s tracking) |
| `/root/scripts/gpu-saturation-watchdog.py` | pi (.24) | Auto-restart stuck llama-server |
| `/root/dashboard/gpu-fleet.html` | pi (.24) | Live HTML dashboard |
| `/etc/systemd/system/llama-server.service` | .8, .110 | llama-server daemons (Nvidia GPUs) |
| `/etc/systemd/system/strix-server.service` | .15 (amdpve) | llama-server daemon (Vulkan, Strix Halo) running unsloth/Qwen3.6-35B-A3B-MTP-GGUF. Note: `llama-server.service` and `llama-server@.service` are **masked** on .15 to prevent port 8080 collisions. |

## Prometheus & Grafana

| Component | URL | Details |
|-----------|-----|---------|
| Grafana | `http://192.168.68.116:3001/` | admin / vault (`GRAFANA_ADMIN_PASSWORD`) |
| GPU Dashboard | `http://192.168.68.116:3001/d/gpu-fleet` | Gauges + time series |
| Prometheus | `http://192.168.68.116:9090/` (internal) | 5 scrape targets |
| GPU Exporters | `:9400/metrics` on .8, .110, .15 | NVIDIA/AMD GPU metrics |
| Router Exporter | `:9401/metrics` on .24 | Router + LiteLLM metrics |

## Known Issues & Watch Points

- **Router startup race**: Compose router.py doesn't call load_roster(). Reload thread sleeps 30s first.
  Fix: trigger roster reload via SSH after restart, or rebuild image with startup load_roster().
- **LiteLLM /metrics**: Requires auth. Prometheus uses `/health/liveliness` as workaround.
- **VRAM (2026-07-15)**: RTX 3090 at ~17/24.6GB (~70%) with **128K context** (reduced from 256K 2026-07-17). RTX 5070 at ~7.8/12.2GB (~65%) with 128K context + MTP. Strix Halo at ~7GB/64GB.
- **RTX 3090 runs `--parallel 2`** with MTP draft (spec-type draft-mtp, spec-draft-n-max 2).
- **RTX 3090 config**: `-c 131072 -ctk turbo4 -ctv turbo4 --parallel 2 --flash-attn on --cont-batching --spec-type draft-mtp`. Context reduced to 128K (2026-07-17, was 256K). VRAM: ~70%. Service: `/home/llmuser/llama-wrapper.sh`.
- **RTX 5070 config (2026-07-15)**: Switched to IQ4_NL + MTP draft (Q8_0) at 128K context. Gen speed: 122 tok/s. VRAM: ~7.8/12.2GB (~65%). Service: `/home/llmuser/llama-wrapper.sh`. Config: `--model gemma-4-12b-it-IQ4_NL.gguf --spec-draft-model gemma-4-12b-it-Q8_0-MTP.gguf --spec-type draft-mtp --spec-draft-n-max 4 --ctx-size 131072`.
- **LiteLLM timeout tuning (verified 2026-07-16 against `/opt/inference-harness/litellm_config.yaml` on CT 116)**: gemma-4-12b 120s, qwen3.6-27B-code 300s, qwen3.6-35B-udq4 300s, strix-moe 300s, syslog-auto routes all 300s. Nginx proxy_read_timeout: 600s. Global request_timeout: 300s.
- **Strix Halo GPU**: Vulkan is the working backend (ROCm/HIP path abandoned — HSA runtime blocked on Debian 13). Build at `/root/llama.cpp/build-vk/`, commit `4fc4ec5` (2026-07-01), ggml 0.15.3 shared-lib arch. Mesa RADV 25.0.7, KHR_coopmat fast path active. ~70 tok/s gen, 532 tok/s prompt. Service: `strix-server.service` on port 8080, model: `qwen3.6-35B-udq4`, alias `strix-moe`, 128K context, flash-attn + q4 KV, multimodal (mmproj loaded).
- **Port conflict detection (2026-07-05)**: All 3 GPU wrappers now detect ghost processes squatting port 8080 before starting. `.8` and `.110` use inline pre-start check in `llama-wrapper.sh`; `.15` uses `/usr/local/bin/port-cleanup.sh` ExecStartPre. Replaces the blanket `pkill -9 -x llama-server` on .15 which would kill ALL llama-server instances regardless of port. Ghost detection was the root cause of .8 crash-looping for 27+ restarts (stale pid 25836 squatting 8080 after OOM kill).
- **Strix Halo thermal safeguard (2026-07-02)**: `strix-server.service` has `-n 8192` (hard generation cap per request). Without it, `--predict` defaults to -1 (infinity) — a runaway request from .123 (Mumuni) decoded 39,868 tokens over 24 min, pushing Tctl to 98°C (crit 89.8°C) and throttling 70→29 t/s. The cap bounds worst-case generation to ~5 min. Do NOT remove `-n` without a replacement ceiling. Sustained load hits ~84°C even at 92s; the APU is fanless/low-flow. Clients MUST also set `max_tokens`.
- **Port 8080 firewall**: amdpve iptables restricts 8080 to 192.168.68.116 (LiteLLM/router host) only. All inbound connections are from .116 (LiteLLM proxied via nginx). Localhost curls hang (SYN dropped). Always test from .116.
- **Router sidecar fallback**: `router.py` `check_gpu_health()` now probes GPU `/health` directly when sidecar at :8090 is absent. Sidecar JSON exporters not deployed on any GPU host — router relies on GPU-direct fallback.
- **Router GPU_MOE_URL bug (fixed 2026-07-01)**: docker-compose had `GPU_MOE_URL=.110:8080` (gemma host) instead of `.15:8080` (amdpve). Corrected.
- **Alert migration**: All alerts now go to `#agent-hub` topics (`alerts-gpu`, `alerts-pm2`, `alerts-infra`) instead of DMs. Cross-agent visibility enabled.
- **tok/s benchmarks**: Measured every 5 min via LiteLLM proxy. Baselines tracked with 30%/50% degradation thresholds.
- **NetBird 502**: Tanko routes through NetBird for litellm.sysloggh.net. Use direct IP if NetBird down.

## GPU Inference Benchmarks (Current)

| GPU | Model | Gen tok/s | Prompt tok/s | Baseline | Context |
|-----|-------|-----------|--------------|----------|---------|
| RTX 3090 (.8) | qwen3.6-27B-code (MTP) | **63** | — | — | **128K** |
| RTX 5070 (.110) | gemma-4-12b (IQ4_NL+MTP) | **191** | — | — | **128K** |
| Strix Halo (.15) | qwen3.6-35B-udq4 | **65** | 140 | — | **128K** |

Benchmarks from 2026-07-17. Strix Halo model: qwen3.6-35B-udq4. RTX 5070 MTP provides 2.7x speedup over pre-upgrade 70 tok/s.
All 3 GPUs now at 128K context (2026-07-17, reduced from 256K for stability).

Benchmarks run through LiteLLM proxy (192.168.68.116:4001) every 5 minutes.
Degradation alerts fire at 30% (warning) and 50% (critical) below baseline.
History stored at `/root/data/toks-history.json` with 7-day rolling window.

**Note (2026-07-01)**: Strix Halo prompt tok/s jumped 209→532 after Vulkan rebuild (cooperative-matrix fast path now active on GFX1151). Baseline may need re-calibration.

## Agent Config Implications (2026-07-15)

### Stable Aliases — CRITICAL

All agent configs MUST use stable role-based aliases, never model-specific names:
- `compression.model: strix-moe` (NOT `qwen3.6-35B-udq4`)
- `auxiliary.vision.model: gpu-light` (NOT `gemma-4-12b`)
- `delegation.model: gpu-dense` (NOT `qwen3.6-27B-code`)
- `auxiliary.web_extract.model: gpu-light`

When the underlying model is swapped, only the LiteLLM config changes — agent configs are untouched.

### Context Windows
- RTX 3090: **128K** (reduced from 256K 2026-07-17) | RTX 5070: **128K** (reduced from 256K) | Strix Halo: **128K**
- **All agents**: 128K ceiling — stable margin. For >128K workloads, use external providers (deepseek)
- Compression threshold 0.65: fires at ~85K (~43K headroom before 128K ceiling)
- Mumuni compression model alias: `strix-moe` with 300s timeout

### Mumuni Agent Profile

Mumuni (CT114, 192.168.68.123) is the primary business assistant. This profile is the reference for all agent configs:

| Setting | Value | Notes |
|---------|-------|-------|
| `model.default` | `syslog-auto` | Weighted pool (55% qwen, 30% strix, 15% gemma) |
| `model.provider` | `custom:litellm` | LiteLLM on CT116 |
| `compression.model` | `strix-moe` | Stable alias — survives model swaps |
| `aux.compression.model` | `strix-moe` | Compression auxiliary model |
| `aux.vision.model` | `gpu-light` | Vision tasks (RTX 5070) |
| `aux.web_extract.model` | `gpu-light` | Web extraction |
| `delegation.model` | `gpu-dense` | Sub-agent reasoning (RTX 3090) |
| `context.max_context_window` | 131072 (128K) | Reduced from 256K 2026-07-17 — stable 128K ceiling |
| `compression.threshold` | 0.65 | Triggers at ~85K |
| `compression.target_ratio` | 0.3 | Compresses to ~38K |
| `compression.protect_last_n` | 40 | Preserves last 40 messages |
| `memory.memory_char_limit` | 800 | Brief memory entries |
| `personalities` | `creative` | Creative assistant personality |
| Platforms | cli, homeassistant, signal, telegram, zulip | All Hermes platforms |
| Main model timeout | 300s | LiteLLM global timeout |
| Compression model timeout | 300s | strix-moe timeout increased from 120s |

### Agent Update Status (2026-07-15)

| Agent | Host | Status |
|-------|------|--------|
| **Mumuni** | CT114 (.123) | ✅ Updated to stable aliases |
| **Tanko** | CT112 (.122) | ✅ Updated to stable aliases |
| **Koby** | CT111 (.129) | ❌ SSH unreachable — needs Zulip DM |
| **Koonimo** | CT113 | ❌ SSH unreachable — needs Zulip DM |
| **Kagenz0** | CT105 | ❌ SSH unreachable — needs Zulip DM |

All Hermes clients MUST set `max_tokens: 4096` — first line of defense before server-side `-n 8192` cap.