infra/assets/ai-optimizer/README.md

# AI Model Optimizer - Manual Skill

**Purpose:** Find optimal ollama configurations for maximum context size and GPU utilization on AMD MI50 GPUs.

**Usage:** Run manually via Hermes skill when needed (not automated).

**Hardware:**
- 2× AMD MI50 GPUs (32GB VRAM each, 64GB total)
- 128GB system RAM
- ROCm: `HSA_OVERRIDE_GFX_VERSION=9.0.6`, `HIP_VISIBLE_DEVICES=0,1`

---

## File Locations

```
STATE:   /opt/data/infra/assets/ai-optimizer/state.json
RESULTS: /opt/data/infra/assets/ai-optimizer/results.csv
REPO:    /opt/data/infra (persistent - do not reclone)
```

---

## Quick Start

```bash
# From Hermes container or any machine with ollama access
ollama-test-model --model devstral-small-2:24b --ctx 65536
```

Or use the full workflow skill for systematic testing.

---

## Model Queues

### GPU Track (Coding - prioritize speed + context on GPU)
1. `deepseek-coder-v2:16b` - Best coding model, fits on GPU
2. `qwen2.5-coder:32b` - Alternative coding model
3. `codellama:34b-instruct` - Legacy option

### RAM Track (Knowledge - prioritize max context)
1. `qwen2.5:72b` - Large knowledge model
2. `nemotron-3-nano:30b` - Efficient large model
3. `mixtral:8x7b-instruct` - MoE architecture

---

## Context Steps (in order)

```
[32768, 65536, 98304, 131072, 163840, 200704, 262144, 327680]
```

---

## Optimization Strategy

### GPU Track (Coding)
- Start: `num_ctx=32768`, `num_gpu=99`, `flash_attn=true`
- Increase context until OOM or tokens/sec < 5
- Record best config before hitting wall
- Target: >10 tokens/sec with max context

### RAM Track (Knowledge)
- Start: `num_ctx=65536`, `num_gpu=50`, `flash_attn=true`
- Allow heavy RAM offload (up to 100GB system RAM)
- Increase context until OOM
- Speed secondary to context size

---

## Manual Testing Workflow

### 1. Quick Model Test

```bash
# Test a model at specific context size
docker exec ollama ollama run <model>:<tag> "Your prompt here"
```

### 2. Check Current State

```bash
cd /opt/data/infra
cat assets/ai-optimizer/state.json
```

### 3. Pull Model (if needed)

```bash
docker exec ollama ollama pull <model>:<tag>
```

### 4. Create Test Modelfile

```bash
docker exec ollama bash -c "cat <<EOF > /root/.ollama/test_${model}.modelfile
FROM ${model}
PARAMETER num_ctx ${num_ctx}
PARAMETER num_gpu ${num_gpu}
PARAMETER flash_attn true
PARAMETER num_predict 4096
PARAMETER num_keep 1024
PARAMETER repeat_penalty 1.1
EOF"

docker exec ollama ollama create test-model -f /root/.ollama/test_${model}.modelfile
```

### 5. Run Benchmark

```bash
# Warm up
docker exec ollama ollama run test-model "Hello" > /dev/null

# Coding prompt
docker exec ollama ollama run test-model "Write a Python async context manager that retries a function with exponential backoff, max 5 retries, and logs each attempt using structlog. Include type hints."

# Knowledge prompt
docker exec ollama ollama run test-model "Explain the complete memory hierarchy in modern GPUs, from registers through L1/L2 caches to VRAM, and how data moves between them during matrix multiplication."
```

### 6. Measure VRAM

```bash
# Try host first
rocm-smi --showmeminfo vram 2>/dev/null || \
# Try via docker
docker exec --privileged ollama rocm-smi --showmeminfo vram 2>/dev/null || \
echo "VRAM unavailable"
```

### 7. Record Results

Update `state.json` and append to `results.csv`:
- tokens/sec from ollama output
- VRAM/RAM usage
- Whether this config is the new best

### 8. Commit Changes

```bash
cd /opt/data/infra
git add assets/ai-optimizer/
git commit -m "ai-optimizer: tested ${model} at ${num_ctx} ctx - ${status}"
git push
```

---

## State File Structure

```json
{
  "track": "gpu",
  "current_model": "deepseek-coder-v2:16b",
  "model_index": 0,
  "phase": "context_scaling",
  "backend": "ollama",
  "current_config": {
    "num_ctx": 32768,
    "num_gpu": 99,
    "flash_attn": true
  },
  "best_configs": {
    "gpu": {},
    "ram": {}
  },
  "completed_models": [],
  "gpu_queue": ["deepseek-coder-v2:16b", "qwen2.5-coder:32b", "codellama:34b-instruct"],
  "ram_queue": ["qwen2.5:72b", "nemotron-3-nano:30b", "mixtral:8x7b-instruct"],
  "context_steps": [32768, 65536, 98304, 131072, 163840, 200704, 262144, 327680],
  "last_updated": "2026-04-30T00:00:00Z"
}
```

---

## Results CSV Format

```csv
timestamp,track,model,backend,phase,num_ctx,num_gpu,flash_attn,tokens_per_sec,vram_gb,ram_gb,status,is_best
2026-04-30T00:00:00Z,gpu,deepseek-coder-v2:16b,ollama,context_scaling,65536,99,true,15.2,52.1,18.4,success,false
```

---

## Skill Usage

Once PR #1 (ai-worker-restricted-access) is merged:

```bash
# From Hermes container, SSH to host for direct ollama access
ssh -i /path/to/key ai-worker@host docker exec ollama ollama run <model>

# Or run the skill directly
ollama-benchmark --model deepseek-coder-v2:16b --track gpu
```

---

## Notes

- **Manual execution only** - No cron job, run when needed
- **Two tracks**: Complete GPU track first (coding models), then RAM track
- **Backend**: ollama (llama.cpp optional for advanced users)
- **Host access**: Use docker exec or SSH for rocm-smi
- **Commit results**: Push best configs to repo for reference