fix: correct quant to UD-Q3_K_XL (Q4 too large for 24GB VRAM)

gpu-dense on RTX 3090 swapped from qwen3.6-27B-code to SmartCode-Fable-5-CoT-Reasoning-QKVO-Qwen-3.6-27B-Distilled (UD-Q4_K_XL, ~17.9GB).\n\nImprovements:\n- ~50% fewer thinking tokens via ThinkingCap finetune\n- Fable 5 CoT distillation for improved coding reasoning\n- Same 27B base, fits RTX 3090 at ~73% VRAM\n- Recommended samplers: temp 0.9, top-p 0.95, top-k 60\n- Context: 128K (fleet standard)

gpu-fleet.prose.md

@@ -14,6 +14,10 @@ description: >
   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%.
+  UPDATED 2026-07-27: gpu-dense swapped to SmartCode-Fable-5-CoT-Reasoning-QKVO-Qwen-3.6-27B-Distilled
+  (UD-Q3_K_XL, ~14.7GB — Q4 was too large for 24GB VRAM with 128K KV cache). ~50% fewer thinking
+  tokens via ThinkingCap finetune + Fable 5 CoT distillation for improved coding reasoning.
+  VRAM ~22.4/24.6GB (91%).
 agent: abiba
 triggers:
   - on model add/remove
@@ -86,7 +90,7 @@ When a model is swapped on a GPU, ONLY the infrastructure layer changes — agen
 | 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-dense` | RTX 3090 (.8) | SmartCode-Fable-5-27B-UD-Q4_K_XL | 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
@@ -96,7 +100,7 @@ but are deprecated for agent configs. Only the stable aliases survive model swap
 
 | 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 |
+| SmartCode-Fable-5-27B-UD-Q3_K_XL | RTX 3090 | .8 (llm-gpu) | ~22.4/24.6GB (91%) | **128K** | q4_0 | 1 | 2048/1024 | ✅ healthy |
 | 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 |
 
@@ -106,7 +110,7 @@ but are deprecated for agent configs. Only the stable aliases survive model swap
 
 | Model | GPU | Weight | RPM Cap | Timeout |
 |-------|-----|--------|---------|---------|
-| qwen3.6-27B-code | RTX 3090 (.8:8080) | **0.55** | 500 | **300s** |
+| SmartCode-Fable-5-27B-UD-Q3_K_XL | 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** |
 
@@ -117,7 +121,7 @@ Note: All syslog-auto entries route directly to GPUs with `api_key: not-needed`.
 | Model | RPM Cap | Notes |
 |-------|---------|-------|
 | strix-moe (qwen3.6-35B-udq4) | 40 | Tight cap — prevents Strix overload |
-| qwen3.6-27B-code | 500 | High cap — primary workhorse |
+| SmartCode-Fable-5-27B-UD-Q3_K_XL | 500 | High cap — primary workhorse (replaces qwen3.6-27B-code) |
 | gemma-4-12b | 500 | High cap — IQ4_NL+MTP, 122 tok/s |
 
 ### Stable Aliases (for agent configs — never change)
@@ -247,10 +251,9 @@ If no SSH access, send Zulip DM via abiba-bot with vault update instructions.
   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 (2026-07-27)**: Swapped to SmartCode-Fable-5-27B-UD-Q3_K_XL (14.7GB). Q4 was too large for 24GB VRAM with 128K context + KV cache overhead. Q3 fits at ~22.4GB (91%). Uses standard llama.cpp build b9190 (turboquant b9150 incompatible with qwen3_5 arch). Config: `-c 131072 -ctk q4_0 -ctv q4_0 --flash-attn on --cont-batching`. Sampler: `--temp 0.9 --top-p 0.95 --top-k 60 --min-p 0.0 --repeat-penalty 1.0`. Service: `/home/llmuser/llama-fable-wrapper.sh`.
-- **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.
+- **LiteLLM timeout tuning (verified 2026-07-27)**: SmartCode-Fable-5-27B 300s, gemma-4-12b 120s, qwen3.6-27B-code 300s (legacy), 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`.
@@ -265,7 +268,7 @@ If no SSH access, send Zulip DM via abiba-bot with vault update instructions.
 
 | GPU | Model | Gen tok/s | Prompt tok/s | Baseline | Context |
 |-----|-------|-----------|--------------|----------|---------|
-| RTX 3090 (.8) | qwen3.6-27B-code (MTP) | **63** | — | — | **128K** |
+| RTX 3090 (.8) | SmartCode-Fable-5-27B-UD-Q3_K_XL | **TBD** | — | — | **128K** |
 | RTX 5070 (.110) | gemma-4-12b (IQ4_NL+MTP) | **191** | — | — | **128K** |
 | Strix Halo (.15) | qwen3.6-35B-udq4 | **65** | 140 | — | **128K** |
 

infrastructure-control.prose.md

@@ -13,10 +13,10 @@ description: >
   against the live system. Policy fields are authoritative. See the
   `verify-before-mutate` skill.
 
-  **Last verified:** 2026-07-26 — CT 114 (mumuni) destroyed. Mumuni moved inside
+  **Last verified:** 2026-07-27 — gpu-dense swapped to SmartCode-Fable-5
-  CT 100 (abiba) — Zulip connected on .24. Verification protocol run: CT 118 set
+  (Qwen3.6-27B distilled, ~50% fewer thinking tokens, improved coding reasoning).
-  to static IP .20, llama-server on .110 restored. All PVE nodes, CT hostnames,
+  Model pricing reduced 100x across all models ($0.15/$0.60 per 1M tokens).
-  and public endpoints confirmed. See data/learnings.md.
+  All LiteLLM models set to 128K max_model_tokens. See data/learnings.md.
 ---
 
 # Infrastructure Control Pattern