abiba-bot/prose-contracts
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: GPU context 256K→128K fleet-wide + Genesis Hermes V3 on Strix Halo

Browse Source 
GPU Changes:
- All 3 GPUs reduced from 256K (-c 262144) to 128K (-c 131072) for stability
- Observed instability near 100K at 256K — 128K is the stable ceiling
- VRAM improved: RTX 3090 ~70% (was 90%), RTX 5070 ~65% (was 88%)
- Strix Halo swapped to LuffyTheFox/Genesis Hermes V3 APEX
  - Hermes agent fine-tune, tensor repair (3 SSM layers, 76% W1 improvement)
  - Uncensored (0/465 refusals), multimodal (mmproj F16)
  - Speed: 65 tok/s gen, 140 tok/s prompt
  - Alias strix-moe maintained

Agent Updates:
- Mumuni: max_context_window 262144→131072, already aligned on strix-moe/0.65
- Tanko: max_context_window 262144→131072
- Koonimo: max_context_window + context_length 262144→131072
- CT114 SSH access confirmed (was 'Zulip only')

LiteLLM (CT116):
- Updated backend model references qwen3.6-35B-udq4→strix-moe
- Removed stale ornith-1.0-35b from model_cost
- Fallback chains updated

Contracts Updated:
- gpu-fleet.prose.md: topology, VRAM, benchmarks, config lines, model assignments
- gpu-self-heal.prose.md: Rule 9/10 context targets
- hermes-config-template.prose.md: template values, Rules 7-9, compression thresholds
- inference-optimization.prose.md: added to repo, 128K recommendation

Compression: 0.65 fires at ~85K (~43K headroom before 128K ceiling)
For >128K workloads: route to external providers (deepseek)
This commit is contained in:
root
2026-07-17 10:59:45 +00:00
parent 65c99dab50
commit b9149bce47
4 changed files with 153 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files
gpu-fleet.prose.md
+27 -23
View File
@@ -9,8 +9,12 @@ 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).
RTX 5070 context: 131K → 256K. VRAM: 88% (10.8/12.2GB).
Compression timeout: 300s (was 120s). Mumuni context: 128K (was 256K).
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,
Hermes agent fine-tune, tensor repair, multimodal with mmproj).
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%.
agent: abiba
triggers:
- on model add/remove
@@ -67,7 +71,7 @@ triggers:
│ CT 8 │ │ CT 110 │ │ CT 15 │ │ pi (.24) │
│ RTX 3090 │ │ RTX 5070 │ │ Strix Halo│ │ GPU Monitor │
│ 24GB │ │ 12GB │ │ 64GB UMA │ │ :9100 │
│ 256K ctx │ │ 256K ctx │ │ 256K ctx │ │ Watchdog │
│ 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 │
@@ -93,9 +97,9 @@ 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) | 22.2/24.6GB (90%) | **256K** 🚀 | turbo4 | 2 | default | ✅ 63 tok/s |
| gemma-4-12b | RTX 5070 | .110 (ocu-llm) | 10.0/12.2GB (82%) | 256K | q4_0 | 2 | 2048/1024 | ✅ healthy |
| qwen3.6-35B-udq4 | Strix Halo Vulkan | .15 (amdpve) | ~9GB/64GB | 256K | q8_0 | 2 | 2048/512 | ✅ healthy |
| 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 |
| Genesis Hermes V3 APEX | Strix Halo Vulkan | .15 (amdpve) | ~10GB/64GB | 128K | q4_0 | 1 | 4096/1024 | ✅ 65 tok/s |
## Routing Configuration (LiteLLM — July 2026)
@@ -104,7 +108,7 @@ 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** |
| qwen3.6-35B-udq4 | Strix Halo (.15:8080) | **0.30** | 60 | **300s** |
| Genesis Hermes V3 APEX | 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.
@@ -113,7 +117,7 @@ Note: All syslog-auto entries route directly to GPUs with `api_key: not-needed`.
| Model | RPM Cap | Notes |
|-------|---------|-------|
| qwen3.6-35B-udq4 | 40 | Tight cap — prevents Strix overload |
| 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 |
@@ -128,11 +132,11 @@ Note: All syslog-auto entries route directly to GPUs with `api_key: not-needed`.
### Fallback Chains
- gemma → qwen
- qwen → gemma
- qwen3.6-35B-udq4 → qwen → gemma
- strix-moe → qwen → gemma
- syslog-auto → qwen → gemma → qwen3.6-35B-udq4
### Why Strix Halo RPM Is Capped
- Direct (qwen3.6-35B-udq4): 40 RPM (tight) — Strix Halo is shared with compression tasks
- 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
@@ -243,12 +247,12 @@ If no SSH access, send Zulip DM via abiba-bot with vault update instructions.
- **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 22.2/24.6GB (90%) with **256K context** (corrected from 131K). RTX 5070 at 10.8/12.2GB (88%) with 256K context + MTP. Strix Halo at ~9GB/64GB.
- **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 262144 -ctk turbo4 -ctv turbo4 --parallel 2 --flash-attn on --cont-batching --spec-type draft-mtp`. Context corrected to 256K (2026-07-15). VRAM: 90%. Service: `/home/llmuser/llama-wrapper.sh`.
- **RTX 5070 config (2026-07-15)**: Switched to IQ4_NL + MTP draft (Q8_0) at 256K context. Gen speed: 122 tok/s (was 70). VRAM: 10.8/12.2GB (88%). No draft model pre-upgrade due to VRAM constraints. 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 262144`.
- **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 (was `ornith-server.service`), model changed to `unsloth/Qwen3.6-35B-A3B-MTP-GGUF` (UD-Q4_K_M), alias `qwen3.6-35B-udq4`, 256K context, flash-attn + q8 KV. MTP support enabled for 1.4-2.2x faster inference.
- **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).
- **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.
@@ -262,12 +266,12 @@ 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** | — | — | **256K** |
| RTX 5070 (.110) | gemma-4-12b (IQ4_NL+MTP) | **191** | — | — | **256K** |
| Strix Halo (.15) | qwen3.6-35B-udq4 | **71** | — | — | **256K** |
| RTX 3090 (.8) | qwen3.6-27B-code (MTP) | **63** | — | — | **128K** |
| RTX 5070 (.110) | gemma-4-12b (IQ4_NL+MTP) | **191** | — | — | **128K** |
| Strix Halo (.15) | Genesis Hermes V3 APEX | **65** | 140 | — | **128K** |
Benchmarks from 2026-07-15 verification run. RTX 5070 MTP provides 2.7x speedup over pre-upgrade 70 tok/s.
All 3 GPUs now at 256K context (2026-07-15).
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.
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.
@@ -288,9 +292,9 @@ All agent configs MUST use stable role-based aliases, never model-specific names
When the underlying model is swapped, only the LiteLLM config changes — agent configs are untouched.
### Context Windows
- RTX 3090: **256K** (was 131K, bumped 2026-07-15) | RTX 5070: **256K** (up from 131K) | Strix Halo: **256K**
- **Mumuni compression context**: 128K (down from 256K) — ensures compression model doesn't timeout
- Compression threshold 0.65: fires at ~85K for 128K context window
- 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
@@ -306,7 +310,7 @@ Mumuni (CT114, 192.168.68.123) is the primary business assistant. This profile i
| `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` | 262144 (256K) | Fixed 2026-07-16 (was 131072 — caused premature compression, WAL #1300) |
| `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 |