gpu-fleet: July 8 optimization sweep — parallel 2 fleet-wide, 128K ctx on NVIDIA, LiteLLM timeout fixes

- All GPUs: --parallel 1 → 2 (6 concurrent slots, was 3)
- .8 RTX 3090: ctx 256K→128K, VRAM 96%→83%, turbo4 KV cache
- .110 RTX 5070: ctx 256K→128K, ubatch 4096→512 (was inverted), VRAM 90%→77%, q4_0 KV
- .15 Strix Halo: parallel 1→2, 256K ctx (41GB free), q8_0 KV, AMD metrics via /sys/class/drm
- LiteLLM: gemma timeout 25→120s, qwen timeout 40→90s, syslog-auto (qwen) 40→90s
- Agent configs: context_length 262144 for syslog-auto, 131072 for direct qwen/gemma
- Updated health-check operation, agent config implications, benchmark table (fixed model↔GPU mapping)

gpu-fleet.prose.md

@@ -5,8 +5,9 @@ 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.
-  Current as of 2026-06-30 post-migration: nginx routes /v1 → LiteLLM,
+  Current as of 2026-07-08: context reduced to 128K on NVIDIA GPUs, parallel 2
-  router runs behind LiteLLM on :9000 (internal-only).
+  on all GPUs, LiteLLM timeouts tuned (gemma 25→120s, qwen 40→90s), router fully
+  deprecated — nginx routes /v1 → LiteLLM directly.
 agent: abiba
 triggers:
   - on model add/remove
@@ -63,21 +64,21 @@ triggers:
 │ CT 8     │  │ CT 110      │  │ CT 15     │  │ pi (.24)     │
 │ RTX 3090 │  │ RTX 5070    │  │ Strix Halo│  │ GPU Monitor  │
 │ 24GB     │  │ 12GB        │  │ 64GB UMA  │  │ :9100        │
-│          │  │             │  │ llama-srv │  │ Watchdog     │
+│ 128K ctx │  │ 128K ctx    │  │ 256K ctx  │  │ Watchdog     │
 │ qwen3.6  │  │ gemma-4-12b │  │ ornith35B │  │ Prometheus   │
-│ 27B-code │  │ :8090       │  │ :8080     │  │ exporter     │
+│ 27B-code │  │ :8080       │  │ :8080     │  │ exporter     │
-│ :8090    │  │ :9400 (exp) │  │ :9400(exp)│  │ :9401        │
+│ :8080    │  │ :9400 (exp) │  │ :9400(exp)│  │ :9401        │
 │ :9400    │  └─────────────┘  └───────────┘  └──────────────┘
 └──────────┘
 ```
 
-## Current Model Assignments
+## Current Model Assignments (2026-07-08)
 
-| Model | GPU | Host | VRAM | Status |
+| Model | GPU | Host | VRAM | Ctx | KV Cache | Parallel | Batch/Ubatch | Status |
-|-------|-----|------|------|--------|
+|-------|-----|------|------|-----|----------|----------|-------------|--------|
-| qwen3.6-27B-code | RTX 3090 | .8 (llm-gpu) | 23.4/24GB | ✅ healthy |
+| qwen3.6-27B-code | RTX 3090 | .8 (llm-gpu) | 20.3/24GB (83%) | 128K | turbo4 | 2 | 512/512 | ✅ healthy |
-| gemma-4-12b | RTX 5070 | .110 (ocu-llm) | 11.2/12.2GB | ✅ healthy |
+| gemma-4-12b | RTX 5070 | .110 (ocu-llm) | 9.4/12.2GB (77%) | 128K | q4_0 | 2 | 2048/512 | ✅ healthy |
-| ornith-1.0-35b | Strix Halo Vulkan | .15 (amdpve) | 22.7/64GB | ✅ healthy |
+| ornith-1.0-35b | Strix Halo Vulkan | .15 (amdpve) | 24.4/64GB (35%) | 256K | q8_0 | 2 | 2048/512 | ✅ healthy |
 
 ## Operations
 
@@ -121,7 +122,19 @@ triggers:
 7. Document keys in knowledge graph
 
 ### list
-Show full fleet status: GPUs, models, VRAM, active requests, circuit breakers, keys
+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: 90s, ornith-1.0-35b: 120s
+   - 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 2026-06-30)
 
@@ -168,9 +181,11 @@ If no SSH access, send Zulip DM via abiba-bot.
 - **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**: RTX 3090 at ~9.6/24GB (60% free), RTX 5070 at ~6.5/12GB (46% free). Healthy.
+- **VRAM (2026-07-08)**: RTX 3090 at 20.3/24GB (83%), RTX 5070 at 9.4/12.2GB (77%), Strix Halo at 24.4/64GB (35%). Context reduced from 256K→128K on NVIDIA GPUs freed ~3.3GB (.8) and ~1.5GB (.110).
-- **RTX 3090 optimization**: `--parallel 2 --batch-size 4096` — doubled concurrent throughput.
+- **All GPUs at `--parallel 2` (2026-07-08)**: Fleet serves 6 concurrent requests (was 3). 2× throughput.
-- **RTX 5070 optimization**: Removed draft model freed ~2GB VRAM for vision tasks. `--parallel 2`.
+- **RTX 3090 config**: `-c 131072 -ctk turbo4 -ctv turbo4 --parallel 2`. No explicit batch flags (512/512 default). Service: `/home/llmuser/llama-wrapper.sh`.
+- **RTX 5070 config**: `--ctx-size 131072 --cache-type-k q4_0 --cache-type-v q4_0 --batch-size 2048 --ubatch-size 512 --parallel 2`. Ubatch fixed 4096→512 (was inverted — ubatch > batch killed prompt throughput). Service: `/home/llmuser/llama-wrapper.sh`.
+- **LiteLLM timeout tuning (2026-07-08)**: gemma-4-12b 25→120s, qwen3.6-27B-code 40→90s, syslog-auto (qwen route) 40→90s. Nginx proxy_read_timeout: 600s. Global request_timeout: 300s. Config at `/opt/inference-harness/litellm_config.yaml`.
 - **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: `ornith-server.service` on port 8080, 256K context, flash-attn + q8 KV.
 - **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)**: `ornith-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`.
@@ -185,8 +200,8 @@ If no SSH access, send Zulip DM via abiba-bot.
 
 | GPU | Model | Gen tok/s | Prompt tok/s | Baseline | Samples |
 |-----|-------|-----------|--------------|----------|---------|
-| RTX 3090 (.8) | gemma-4-12b | 75 | 305 | 74 | 6 |
+| RTX 3090 (.8) | qwen3.6-27B-code | 75 | 305 | 74 | 6 |
-| RTX 5070 (.110) | qwen3.6-27B-code | 75 | 323 | 75 | 6 |
+| RTX 5070 (.110) | gemma-4-12b | 75 | 323 | 75 | 6 |
 | Strix Halo (.15) | ornith-1.0-35b | 70 | 532 | 70 | 6 |
 
 Benchmarks run through LiteLLM proxy (192.168.68.116:4001) every 5 minutes.
@@ -194,3 +209,13 @@ 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 (2026-07-08)
+
+With NVIDIA GPUs at 128K context:
+- Agents using `syslog-auto` (50/50 qwen+ornith): keep `context_length: 262144` — ornith supports it, Litellm fallbacks handle qwen overflow
+- Agents using `qwen3.6-27B-code` directly: set `context_length: 131072` and `max_tokens: 4096` per thermal safety rule
+- Agents using `gemma-4-12b` directly (auxiliary tasks): set `context_length: 131072`
+- Compression threshold at 0.65: fires at ~170K for syslog-auto (262K ctx), ~85K for direct qwen/gemma (128K ctx)
+- All Hermes clients MUST set `max_tokens: 4096` — first line of defense before server-side `-n 8192` cap
+- Port 8080 is used on all 3 GPU hosts (not 8090 as previously documented)

hermes-agent-baseline.prose.md

@@ -5,7 +5,7 @@ version: 1.0.0
 description: >
   Canonical known-good baseline for all Syslog Hermes agents. Captures the exact
   configuration state, keys, workarounds, and audit procedure. When an agent's
-  configuration goes sideways, restore from this baseline. Last verified 2026-07-05.
+  configuration goes sideways, restore from this baseline. Last verified 2026-07-08. GPU context reduced to 128K on .8/.110, parallel 2 fleet-wide.
 author: Abiba (pi agent)
 ---
 
@@ -22,12 +22,12 @@ done
 
 ## Agent Map
 
-| Agent | CT | Node | IP | LiteLLM Key | LiteLLM Alias |
+| Agent | CT | Node | IP | LiteLLM Key | LiteLLM Alias | Platform |
-|-------|-----|------|-----|-------------|---------------|
+|-------|-----|------|-----|-------------|---------------|----------|
-| Tanko | 112 | amdpve | .122 | `sk-CggiHWlamQyShxWC3Hx6uw` | `tanko` |
+| Tanko | 112 | amdpve | .122 | `sk-CggiHWlamQyShxWC3Hx6uw` | `tanko` | Hermes |
-| Mumuni | 114 | minipve | .123 | `sk-VrqCNlwUgzoNGOpikJ7nwQ` | `mumuni` |
+| Mumuni | 114 | minipve | .123 | `sk-VrqCNlwUgzoNGOpikJ7nwQ` | `mumuni` | Hermes |
-| Tdunna | 111 | amdpve | ? | `sk-6sbCNjz2T6lTVDBdlNHXsA` | `tdunna` |
+| Tdunna | 111 | amdpve | srv1079750 | `sk-Qvzi4uYQBhlSK_XstEhcyQ` | `tdunna` | **pi** |
-| Baggy | 113 | amdpve | ? | `sk-krnw_zGBwvvL5b7l2t-s-A` | `baggy` |
+| Baggy | 113 | amdpve | ? | `sk-krnw_zGBwvvL5b7l2t-s-A` | `baggy` | Hermes |
 
 Access: `pct-run <CT_ID> <command>` — no IPs needed. GPU hosts (.8, .110, .15) use SSH.
 
@@ -45,6 +45,8 @@ Agent (systemd) → LITELLM_API_KEY → LiteLLM (:116/v1) → Router (:9000) →
 
 ## Config Pattern — Mandatory Fields
 
+### For Hermes Agents (Tanko, Mumuni, Baggy)
+
 Every agent's `/root/.hermes/config.yaml` (or `/home/jerome/.hermes/config.yaml`) MUST have:
 
 ### 1. Main Model
@@ -154,6 +156,51 @@ pct-run <CT> grep -A8 "vision:" /root/.hermes/config.yaml | grep api_key
 # Must show both api_key: sk-... and api_key_env: LITELLM_API_KEY
 ```
 
+### For pi Agents (Tdunna)
+
+Tdunna (CT111) runs pi 0.80.3 via PM2 with the Zulip extension (router-worker architecture).
+Config files: `~/.pi/agent/models.json`, `~/.pi/agent/settings.json`.
+
+**models.json** — Must only list models authorized for the agent's LiteLLM key:
+```json
+{
+  "providers": {
+    "syslog-harness": {
+      "baseUrl": "http://192.168.68.116/v1",
+      "api": "openai-completions",
+      "apiKey": "sk-...",
+      "models": [
+        { "id": "syslog-auto" },
+        { "id": "ornith-1.0-35b" },
+        { "id": "qwen3.6-27B-code" },
+        { "id": "gemma-4-12b" }
+      ]
+    }
+  }
+}
+```
+
+**settings.json** — Always use `syslog-auto` as default:
+```json
+{
+  "defaultProvider": "syslog-harness",
+  "defaultModel": "syslog-auto"
+}
+```
+
+**Validation**: Verify models match LiteLLM's authorized list:
+```bash
+curl -s http://192.168.68.116:4000/v1/models \
+  -H "Authorization: Bearer $(grep apiKey ~/.pi/agent/models.json | head -1 | cut -d'"' -f4)" \
+  | jq '.data[].id'
+```
+
+**Stuck worker detection**: In PM2 logs, `workers=[<id>:busy:N]` with growing N indicates
+a stuck worker (model error, no `agent_end` emitted). Fix: correct models.json, delete
+stale sessions from `~/.pi/agent/sessions/zulip/`, restart PM2.
+
+**Service**: `pm2 restart koby-zulip`, health at `:9201/health`.
+
 ## Systemd Pattern
 
 For agents where the gateway runs as a user service:
@@ -210,5 +257,6 @@ If they differ → ghost detected → kill ghost → start fresh.
 
 | Date | Change |
 |------|--------|
+| 2026-07-08 | Tdunna: fixed model mismatch (qwen3.6-35B-A3B→syslog-auto), added pi-specific config section. Key updated to sk-Qvzi4uYQBhlSK_XstEhcyQ. Added Failure Mode #11 to zulip-adapter-lessons. |
 | 2026-07-06 | Port conflict detection added to all 3 GPU wrappers. Consolidated health check script deployed. Zulip streaming edit_message enabled for Tanko/Mumuni. |
 | 2026-07-05 | Baseline created. All 4 agents audited, master key removed, api_key workaround applied |

hermes-config-template.prose.md

@@ -5,7 +5,8 @@ description: >
   Standard Hermes configuration template for Syslog Solution LLC agents.
   Enforces shared infrastructure setup (Firecrawl, SearXNG, local models,
   RA-H OS MCP) while keeping agent-specific API keys and model choices.
-  Updated 2026-06-30: new LiteLLM keys, nginx routing, router back online.
+  Updated 2026-07-08: GPU context reduced to 128K on NVIDIA (.8, .110),
+  parallel 2 on all GPUs, LiteLLM timeouts tuned, context_length guidance added.
 ---
 
 ## Maintains
@@ -92,7 +93,8 @@ model:
   base_url: http://192.168.68.116/v1
   api_key_env: LITELLM_API_KEY    # Set in /etc/environment AND ~/.hermes/.env
   max_tokens: 4096                # ⚠️ CRITICAL: Prevents unbounded generation
-  context_length: 262144          # Must match model's max context window
+  context_length: 262144          # For syslog-auto (ornith route supports 256K).
+                                  # Set 131072 if using qwen3.6-27B-code or gemma-4-12b directly.
 
 fallback_providers:
   provider: deepseek
@@ -123,8 +125,9 @@ compression:
   enabled: true
   model: gemma-4-12b              # ⚠️ Must match auxiliary.compression.model
   provider: harness
-  max_context_window: 262144      # Must match model's actual capacity
+  max_context_window: 262144      # For syslog-auto (ornith supports 256K).
-  threshold: 0.65                 # Fires at ~170K for 262K window (not 0.25!)
+                                  # Set 131072 if using qwen or gemma directly.
+  threshold: 0.65                 # Fires at ~170K for 262K window, ~85K for 128K
   target_ratio: 0.30
   protect_last_n: 40
   hygiene_hard_message_limit: 350
@@ -236,6 +239,28 @@ The following MUST be identical across ALL profiles:
 - `max_context_window: 262144` MUST match the model's actual capacity
 - See `devops-hermes-compression` skill for full reference
 
+### Rule 9: Default Model Must Be `syslog-auto` (All Agents)
+- **Hermes agents**: `model.default: syslog-auto`, `custom_providers[0].model: syslog-auto`
+- **pi agents**: `defaultModel: syslog-auto` in `settings.json`, first model in `models.json`
+- `syslog-auto` is the LiteLLM routing model — it load-balances between ornith-1.0-35b
+  and qwen3.6-27B-code, with gemma-4-12b as fallback. Using it protects against:
+  - Model name typos that cause 403 errors and silent worker failures
+  - Single GPU downtime (routing falls back automatically)
+  - Key/model authorization mismatches
+- **Exception**: Sub-agent profiles (Mumuni's 6 profiles) may specify explicit models
+  for specialized tasks, but MUST validate those models exist in the key's authorized list
+
+### Rule 10: Validate Model IDs Before Deployment (pi Agents)
+- After configuring a pi agent's `models.json`, verify every model ID:
+  ```bash
+  curl -s http://192.168.68.116:4000/v1/models \
+    -H "Authorization: Bearer <AGENT_KEY>" | jq '.data[].id'
+  ```
+- All model IDs in `models.json` MUST appear in the LiteLLM response
+- The agent's API key may have a SUBSET of the full model catalog — check per-key
+- A non-existent model ID causes 403 errors that silently break the pi RPC worker
+  (no `agent_end` emitted, worker stays "busy", Zulip messages pile up unprocessed)
+
 ## Execution
 
 1. **Check current config** — Read the target agent's config.yaml

infrastructure-update.prose.md

@@ -64,7 +64,7 @@ Before ANY update wave:
 **Verify after Wave 2:**
 - All VMs/CTs running: check via Proxmox API
 - LiteLLM healthy: `curl localhost:4000/health/liveliness` (via CT 116)
-- GPU servers responding: check :8080 on VM 101, VM 103; check ornith via router
+- GPU servers responding: check :8080 on VM 101, VM 103; check ornith via router (http://192.168.68.116/health/unified — .15:8080 is firewalled to .116 only)
 - Zulip agents connected: check Mumuni/Tanko gateway state
 - Abiba PM2 processes online: `pm2 status`
 
@@ -120,6 +120,8 @@ Before Wave 1, snapshot these files:
 /opt/home_stack/docker-compose.yml                      (VM 109 .7)
 /opt/audiobookshelf/docker-compose.yml                  (VM 109 .7)
 /root/.pi/agent/extensions/config.yaml                  (CT 100 .24)
+/etc/systemd/system/ornith-server.service               (amdpve .15)
+/etc/systemd/system/llama-server.service                (VM 101 .8, VM 103 .110)
 ```
 
 Run: `mkdir -p /tmp/infra-update-backup-$(date +%Y%m%d) && rsync -av ...`

zulip-adapter-lessons.prose.md

@@ -73,6 +73,25 @@ description: >
 - **Fix**: If edit fails, send the response as a new message instead
 - **Applies to**: pi extension (fixed), Hermes adapter (verify edit fallback exists)
 
+### 11. Model-ID Mismatch Causes Silent Worker Failure (pi-specific)
+- **Symptom**: User sends DM, sees placeholder, but never gets response. Worker stays
+  "busy" forever, accumulating pending replies (10+). Health endpoint shows green but
+  `messages_processed` stalls.
+- **Root cause**: Agent's model config (`models.json` or `settings.json`) references a
+  model ID that doesn't exist in LiteLLM's authorized model list. Example: `qwen3.6-35B-A3B`
+  configured but LiteLLM only exposes `ornith-1.0-35b` under that key. pi's session
+  workers emit 403 on first prompt, then never recover because the error doesn't trigger
+  `agent_end` — worker stays `busy` and all subsequent messages pile up in the steer queue.
+- **Detection**: Compare `~/.pi/agent/models.json` model IDs against `curl -H "Authorization: Bearer <KEY>" http://192.168.68.116/v1/models` output. A stuck worker shows
+  `workers=[<id>:busy:N]` with growing N in extension logs.
+- **Fix**: (1) Update `models.json` to only include models from the authorized list.
+  (2) Set `defaultModel` to `syslog-auto` (safe routing model). (3) Delete stale session
+  JSONL files from `~/.pi/agent/sessions/zulip/`. (4) Restart PM2 process.
+- **Prevention**: Use `syslog-auto` as default model for all agents — it handles model
+  routing and fallback automatically. Direct model IDs (`ornith-1.0-35b`, etc.) should
+  only be used when explicitly requested. Validate model IDs at agent setup time.
+- **Applies to**: pi extension (Tdunna CT111, fixed 2026-07-08), any agent using `syslog-harness` provider
+
 ## Deployment Checklist
 
 When deploying a new Zulip adapter, verify:
@@ -86,3 +105,4 @@ When deploying a new Zulip adapter, verify:
 - [ ] `@all-bots` detected via configurable user_id
 - [ ] Poll uses long-poll (not `dont_block=true` polling)
 - [ ] Stuck/idle detection accounts for quiet periods
+- [ ] Model IDs in config validated against `GET /v1/models` with actual API key