gpu-fleet.prose.md

@@ -9,9 +9,9 @@ description: >
   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 Genesis Hermes V3 APEX (LuffyTheFox, 24GB, uncensored,
+  UPDATED 2026-07-17: Strix Halo model swapped to Genesis Hermes V3 APEX (LuffyTheFox, 24GB, uncensored,
   Hermes agent fine-tune, tensor repair, multimodal with mmproj).
+  RTX 5070 swapped to HauhauCS Gemma4-12B QAT Uncensored Balanced (Q4_K_M, 87 tok/s, 0/465 refusals).
   Instability observed near 100K at 256K. 128K is the stable ceiling.
   For larger context needs → fall back to external providers (deepseek).
   VRAM headroom improved: RTX 3090 ~70%, RTX 5070 ~65%.
@@ -93,14 +93,20 @@ When a model is swapped on a GPU, ONLY the infrastructure layer changes — agen
 **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)
+## Current Model Assignments (2026-07-17)
 
 | 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 |
+| gemma-4-12b (HauhauCS QAT) | RTX 5070 | .110 (ocu-llm) | ~10.0/12.2GB (82%) | 128K | q4_0 | 1 | 2048/1024 | ✅ 87 tok/s |
 | Genesis Hermes V3 APEX | Strix Halo Vulkan | .15 (amdpve) | ~10GB/64GB | 128K | q4_0 | 1 | 4096/1024 | ✅ 65 tok/s |
 
+> **RTX 5070 model swap (2026-07-17)**: Switched from `gemma-4-12b-it-IQ4_NL` (Unsloth, 191 tok/s)
+> to `HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced` (Q4_K_M QAT, 87 tok/s).
+> Trade: 54% slower generation for QAT quality, 0/465 refusals, and agent-optimized tuning.
+> MTP draft also swapped: Q8_0 (444MB) → tuned draft (242MB), saving 200MB VRAM.
+> Role unchanged: gpu-light (vision, web extract, light auxiliary tasks).
+
 ## Routing Configuration (LiteLLM — July 2026)
 
 ### syslog-auto Weighted Pool (Direct GPU — bypasses router)
@@ -119,7 +125,7 @@ Note: All syslog-auto entries route directly to GPUs with `api_key: not-needed`.
 |-------|---------|-------|
 | 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 |
+| gemma-4-12b | 500 | HauhauCS QAT Uncensored Balanced + MTP, 87 tok/s |
 
 ### Stable Aliases (for agent configs — never change)
 
@@ -127,7 +133,7 @@ Note: All syslog-auto entries route directly to GPUs with `api_key: not-needed`.
 |-------|---------|-----------|---------|
 | `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 |
+| `gpu-light` | 500 | RTX 5070 (HauhauCS QAT) | Vision, web extract, light tasks |
 
 ### Fallback Chains
 - gemma → qwen
@@ -249,8 +255,8 @@ If no SSH access, send Zulip DM via abiba-bot with vault update instructions.
 - **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`.
+- **RTX 3090 config**: `-c 131072 -ctk turbo4 -ctv turbo4 --parallel 1 --flash-attn on --cont-batching --spec-type draft-mtp`. Context 128K, single slot for full 128K per-request. VRAM: ~85%. Service: `/home/llmuser/llama-wrapper.sh`.
+- **RTX 5070 config (2026-07-17)**: HauhauCS Gemma4-12B QAT Uncensored Balanced (Q4_K_M) + tuned MTP draft (242MB) at 128K context, single slot. Gen speed: 87 tok/s (vs 191 IQ4_NL). VRAM: ~10.0/12.2GB (~82%). Service: `/home/llmuser/llama-wrapper.sh`. Model: `Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-Q4_K_M.gguf`, MTP: `mtp-gemma-4-12B-it.gguf`, mmproj: `mmproj-Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-BF16.gguf`. Recommended sampling: temp 0.6, top_k 64, top_p 0.9, min_p 0.05, repeat_penalty 1.1.
 - **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: `LuffyTheFox/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V3-GGUF` (APEX quant), alias `strix-moe`, 128K context, flash-attn + q4 KV, multimodal (mmproj loaded). Hermes agent fine-tune, tensor repair (SSM layers fixed via SVD), uncensored (0/465 refusals).
 - **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).
@@ -267,7 +273,7 @@ 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 5070 (.110) | gemma-4-12b (IQ4_NL+MTP) | **191** | — | — | **128K** |
+| RTX 5070 (.110) | HauhauCS QAT Uncensored Balanced | **87** | — | — | **128K** |
 | Strix Halo (.15) | Genesis Hermes V3 APEX | **65** | 140 | — | **128K** |
 
 Benchmarks from 2026-07-17. Strix Halo swapped to Genesis Hermes V3 APEX (LuffyTheFox). RTX 5070 MTP provides 2.7x speedup over pre-upgrade 70 tok/s.