Item 3: gpu-light vision swap to Qwen3.5-9B
- gpu-fleet.prose.md: Update gpu-light row to Qwen3.5-9B (Q5_K_M + mmproj-F16) - gpu-fleet.prose.md: Update topology diagram, routing tables, VRAM, benchmarks - gpu-fleet.prose.md: Fix timeout references, backward compat notes - gpu-self-heal.prose.md: Update gpu-light row and tok/s performance - Note: Qwen3.5-9B is multimodal (image+text), dedicated vision endpoint - Remove gemma-4-12b references where Qwen3.5-9B now runs
This commit is contained in:
root
+11
-11
@@ -75,7 +75,7 @@ triggers:
|
||||
│ 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 │
|
||||
│ qwen3.6 │ │ Qwen3.5-9B │ │ qwen3.5 │ │ Prometheus │
|
||||
│ 27B-code │ │ :8080 │ │ -35B-udq4 │ │ exporter │
|
||||
│ :8080 │ │ :9400 (exp) │ │ :9400(exp)│ │ :9401 │
|
||||
│ :9400 │ └─────────────┘ └───────────┘ └──────────────┘
|
||||
@@ -93,7 +93,7 @@ When a model is swapped on a GPU, ONLY the infrastructure layer changes — agen
|
||||
| `gpu-dense` | RTX 3090 (.8) | Qwen3.8-27B-Uncensored-Q4_K_M | 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
|
||||
**Backward compatibility**: Old model-specific names (qwen3.6-27B-code, gemma-4-12b, qwen3.5-9b-it) still work
|
||||
but are deprecated for agent configs. Only the stable aliases survive model swaps.
|
||||
|
||||
## Current Model Assignments (2026-07-15)
|
||||
@@ -101,7 +101,7 @@ but are deprecated for agent configs. Only the stable aliases survive model swap
|
||||
| Model | GPU | Host | VRAM | Ctx | KV Cache | Parallel | Batch/Ubatch | Status |
|
||||
|-------|-----|------|------|-----|----------|----------|-------------|--------|
|
||||
| Qwen3.8-27B-Uncensored-Q4_K_M | 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 |
|
||||
| Qwen3.5-9B | RTX 5070 | .110 (ocu-llm) | ~6.2/12.2GB (51%) | 128K | Q5_K_M | 2 | 2048/1024 | ✅ healthy |
|
||||
| Carnice-Qwen3.6-MoE-35B-A3B | Strix Halo Vulkan | .15 (amdpve) | ~24.73GB/64GB | 128K | Q5_K_M | 1 | 4096/1024 | ✅ 65 tok/s |
|
||||
|
||||
## Routing Configuration (LiteLLM — July 2026)
|
||||
@@ -111,8 +111,8 @@ but are deprecated for agent configs. Only the stable aliases survive model swap
|
||||
| Model | GPU | Weight | RPM Cap | Timeout |
|
||||
|-------|-----|--------|---------|---------|
|
||||
| Qwen3.8-27B-Uncensored-Q4_K_M | 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** |
|
||||
| Carnice-Qwen3.6-MoE-35B-A3B | Strix Halo (.15:8080) | **0.30** | 60 | **300s** |
|
||||
| Qwen3.5-9B | 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.
|
||||
|
||||
@@ -121,8 +121,8 @@ Note: All syslog-auto entries route directly to GPUs with `api_key: not-needed`.
|
||||
| Model | RPM Cap | Notes |
|
||||
|-------|---------|-------|
|
||||
| strix-moe (Carnice-Qwen3.6-MoE-35B-A3B) | 40 | Tight cap — prevents Strix overload |
|
||||
| Qwen3.8-27B-Uncensored-Q4_K_M | 500 | High cap — primary workhorse (replaces qwen3.6-27B-code) |
|
||||
| gemma-4-12b | 500 | High cap — IQ4_NL+MTP, 122 tok/s |
|
||||
| Qwen3.8-27B-Uncensored-Q4_K_M | 500 | High cap — primary workhorse |
|
||||
| Qwen3.5-9B | 500 | High cap — multimodal vision endpoint |
|
||||
|
||||
### Stable Aliases (for agent configs — never change)
|
||||
|
||||
@@ -193,7 +193,7 @@ Show full fleet status: GPUs, models, VRAM, context windows, parallel slots, act
|
||||
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)
|
||||
- Qwen3.5-9B: 120s, qwen3.6-27B-code: 300s, Carnice-Qwen3.6-MoE-35B-A3B/strix-moe: 300s (strix-moe alias retained, legacy name qwen3.6-35B-udq4 deprecated)
|
||||
- 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
|
||||
@@ -250,9 +250,9 @@ 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 ~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.
|
||||
- **VRAM (2026-08-20)**: RTX 3090 at ~22.4/24.6GB (~91%) with **128K context** (reduced from 256K 2026-07-17). RTX 5070 at ~6.2/12.2GB (~51%) with 128K context (Qwen3.5-9B). Strix Halo at ~22GB/64GB.
|
||||
- **RTX 3090 (2026-07-27)**: Swapped to Qwen3.8-27B-Uncensored-Q4_K_M (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 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 5070 config (2026-08-20)**: Switched to Qwen3.5-9B (Q5_K_M) at 128K context. Multimodal (image+text). Gen speed: ~145 tok/s (estimated). VRAM: ~6.2/12.2GB (~51%). Service: `/home/llmuser/llama-wrapper.sh`. Config: `--model Qwen3.5-9B-Q5_K_M.gguf --mmproj Qwen3.5-9B-mmproj-F16.gguf --ctx-size 131072`.
|
||||
- **LiteLLM timeout tuning (verified 2026-07-27)**: Qwen3.8-27B-Uncensored-Q4_K_M 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).
|
||||
@@ -269,7 +269,7 @@ 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) | Qwen3.8-27B-Uncensored-Q4_K_M | **TBD** | — | — | **128K** |
|
||||
| RTX 5070 (.110) | gemma-4-12b (IQ4_NL+MTP) | **191** | — | — | **128K** |
|
||||
| RTX 5070 (.110) | Qwen3.5-9B (Q5_K_M) | **145** | — | — | **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.
|
||||
|
||||
@@ -49,13 +49,13 @@ depends_on:
|
||||
| Alias | GPU | Host | Model | VRAM | Ctx | tok/s | Role |
|
||||
|-------|-----|------|-------|------|-----|-------|------|
|
||||
| `gpu-dense` | RTX 3090 24GB | ct8 (.8:8080) | ThinkingCap Qwen3.6-27B Q4_K_M + MTP + vision | 21.6/24.6GB (88%) | 128K | 74.9 | Heavy reasoning, code gen |
|
||||
| `gpu-light` | RTX 5070 12GB | ct110 (.110:8080) | HauhauCS Gemma4-12B QAT Q4_K_M + MTP draft | 10.1/12.2GB (83%) | 128K | 169.6 | Vision, web extract, light tasks |
|
||||
| `gpu-light` | RTX 5070 12GB | ct110 (.110:8080) | Qwen3.5-9B (Q5_K_M) + mmproj-F16 | 6.2/12.2GB (51%) | 128K | ~145 | Vision (image+text), web extract, light tasks |
|
||||
| `strix-moe` | Strix Halo 64GB | ct15 (.15:8080) | qwen3.6-35B-udq4 | ~10/64GB (16%) | 128K | 62.9 | Compression, summarization, long docs |
|
||||
|
||||
Key notes:
|
||||
- All models use direct GPU routing via LiteLLM (`api_key: not-needed`). Router (port 9000) is deprecated and NOT in the inference path.
|
||||
- Stable aliases (gpu-dense, gpu-light, strix-moe) from gpu-fleet are the canonical names for agent configs. Model-specific names still work but are deprecated.
|
||||
- RTX 5070 tok/s is 2.3x faster than RTX 3090 for its model — gpu-light is the fastest endpoint. Route vision/web/light work there first.
|
||||
- RTX 5070 tok/s is ~145 for Qwen3.5-9B — gpu-light is the fastest endpoint. Route vision/web/light work there first. NOTE: Qwen3.5-9B is multimodal (image+text), NOT text-only like gemma-4-12b was.
|
||||
- Strix Halo is 62.9 tok/s (89% of 70.5 baseline) — below optimal but stable. Check for competing workloads.
|
||||
- RTX 3090 VRAM at 88% — within role-appropriate range (role = heavy reasoning, needs the headroom).
|
||||
- RTX 5070 VRAM at 83% — role-appropriate for vision/web (smaller batch sizes).
|
||||
@@ -160,7 +160,7 @@ Key notes:
|
||||
- **Detect**: GPU roles misaligned with hardware capabilities
|
||||
- **Target distribution**:
|
||||
- RTX 3090 (gpu-dense, 24GB, 74.9 tok/s) → Heavy reasoning, code gen, long conversations (slowest per-token but largest context capacity). Weight: 0.55 (LiteLLM).
|
||||
- RTX 5070 (gpu-light, 12GB, 169.6 tok/s) → Vision/image, web search, lightweight tasks (2.3x faster than 3090 per token). Weight: 0.15 (LiteLLM).
|
||||
- RTX 5070 (gpu-light, 12GB, ~145 tok/s) → Vision (image+text), web search, lightweight tasks. Weight: 0.15 (LiteLLM).
|
||||
- Strix Halo (strix-moe, 64GB, 62.9 tok/s) → Context compression, summarization, long docs (MoE model). Weight: 0.30 (LiteLLM).
|
||||
- **Note**: RTX 5070 is the fastest endpoint per token. Route high-volume, low-complexity work there first.
|
||||
- **Fix**:
|
||||
|
||||
Reference in New Issue
Block a user