gortium/infra
1
1
Fork 0
Code Issues 13 Pull Requests 21 Actions Packages Projects 1 Releases Wiki Activity

Add AI model optimization cron job #2

Closed
Hermes wants to merge 3 commits from ai-optimizer-cron-job into master
pull from: ai-optimizer-cron-job
merge into: gortium:master
Conversation 0 Commits 3 Files Changed 3 +231
3 changed files with 231 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Showing only changes of commit 0ec198dec2 - Show all commits

135
assets/ai-optimizer/README.md
View File

@@ -1,13 +1,13 @@
# AI Model Optimization Cron Job # AI Model Optimizer - Manual Skill
**Purpose:** Automatically find optimal ollama/llama.cpp configurations for maximum context size and hardware utilization. **Purpose:** Find optimal ollama configurations for maximum context size and GPU utilization on AMD MI50 GPUs.
**Schedule:** Every hour **Usage:** Run manually via Hermes skill when needed (not automated).
**Hardware:** **Hardware:**
- 2× AMD MI50 GPUs (32GB VRAM each, 64GB total) - 2× AMD MI50 GPUs (32GB VRAM each, 64GB total)
- 128GB system RAM - 128GB system RAM
- ROCm: HSA_OVERRIDE_GFX_VERSION=9.0.6, HIP_VISIBLE_DEVICES=0,1 - ROCm: `HSA_OVERRIDE_GFX_VERSION=9.0.6`, `HIP_VISIBLE_DEVICES=0,1`
--- ---
@@ -21,21 +21,33 @@ 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 ## Model Queues
### GPU Track (Coding - prioritize speed + context on GPU) ### GPU Track (Coding - prioritize speed + context on GPU)
1. `devstral-small-2:24b` 1. `deepseek-coder-v2:16b` - Best coding model, fits on GPU
2. `qwen2.5-coder:32b` 2. `qwen2.5-coder:32b` - Alternative coding model
3. `codellama:34b-instruct` 3. `codellama:34b-instruct` - Legacy option
### RAM Track (Knowledge - prioritize max context) ### RAM Track (Knowledge - prioritize max context)
1. `qwen2.5:72b` 1. `qwen2.5:72b` - Large knowledge model
2. `nemotron-3-nano:30b` 2. `nemotron-3-nano:30b` - Efficient large model
3. `mixtral:8x7b-instruct` 3. `mixtral:8x7b-instruct` - MoE architecture
--- ---
## Context Steps (in order) ## Context Steps (in order)
``` ```
[32768, 65536, 98304, 131072, 163840, 200704, 262144, 327680] [32768, 65536, 98304, 131072, 163840, 200704, 262144, 327680]
``` ```
@@ -45,45 +57,49 @@ REPO: /opt/data/infra (persistent - do not reclone)
## Optimization Strategy ## Optimization Strategy
### GPU Track (Coding) ### GPU Track (Coding)
- Start: num_ctx=32768, num_gpu=99, flash_attn=true - Start: `num_ctx=32768`, `num_gpu=99`, `flash_attn=true`
- Increase context until OOM or tokens/sec < 5 - Increase context until OOM or tokens/sec < 5
- Record best config before hitting wall - Record best config before hitting wall
- Target: >10 tokens/sec with max context - Target: >10 tokens/sec with max context
### RAM Track (Knowledge) ### RAM Track (Knowledge)
- Start: num_ctx=65536, num_gpu=50, flash_attn=true - Start: `num_ctx=65536`, `num_gpu=50`, `flash_attn=true`
- Allow heavy RAM offload (up to 100GB system RAM) - Allow heavy RAM offload (up to 100GB system RAM)
- Increase context until OOM - Increase context until OOM
- Speed secondary to context size - Speed secondary to context size
--- ---
## Each Run - Step by Step ## 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
### 1. Read State
```bash ```bash
cd /opt/data/infra cd /opt/data/infra
cat assets/ai-optimizer/state.json cat assets/ai-optimizer/state.json
``` ```
### 2. Determine Next Test
- Read `track` (gpu or ram)
- Get `current_model` from queue at `model_index`
- Get `current_config` for parameters to test
- Select next context step from `context_steps`
### 3. Pull Model (if needed) ### 3. Pull Model (if needed)
```bash ```bash
docker exec ollama ollama list | grep -q "<model>" || docker exec ollama ollama pull <model> docker exec ollama ollama pull <model>:<tag>
``` ```
### 4. Create Test Modelfile ### 4. Create Test Modelfile
```bash ```bash
docker exec ollama bash -c "cat <<EOF > /root/.ollama/test_${model}.modelfile docker exec ollama bash -c "cat <<EOF > /root/.ollama/test_${model}.modelfile
FROM ${model} FROM ${model}
PARAMETER num_ctx ${current_config.num_ctx} PARAMETER num_ctx ${num_ctx}
PARAMETER num_gpu ${current_config.num_gpu} PARAMETER num_gpu ${num_gpu}
PARAMETER flash_attn ${current_config.flash_attn} PARAMETER flash_attn true
PARAMETER num_predict 4096 PARAMETER num_predict 4096
PARAMETER num_keep 1024 PARAMETER num_keep 1024
PARAMETER repeat_penalty 1.1 PARAMETER repeat_penalty 1.1
@@ -93,6 +109,7 @@ docker exec ollama ollama create test-model -f /root/.ollama/test_${model}.model
``` ```
### 5. Run Benchmark ### 5. Run Benchmark
```bash ```bash
# Warm up # Warm up
docker exec ollama ollama run test-model "Hello" > /dev/null docker exec ollama ollama run test-model "Hello" > /dev/null
@@ -104,7 +121,8 @@ docker exec ollama ollama run test-model "Write a Python async context manager t
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." 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 (if possible) ### 6. Measure VRAM
```bash ```bash
# Try host first # Try host first
rocm-smi --showmeminfo vram 2>/dev/null || \ rocm-smi --showmeminfo vram 2>/dev/null || \
@@ -114,24 +132,14 @@ echo "VRAM unavailable"
``` ```
### 7. Record Results ### 7. Record Results
- Parse tokens/sec from ollama output
- Record VRAM/RAM usage
- Update `best_configs` if improved
### 8. Update State Update `state.json` and append to `results.csv`:
```python - tokens/sec from ollama output
if test_successful: - VRAM/RAM usage
if context_step < max_reached: - Whether this config is the new best
current_config.num_ctx = next_context_step
else: ### 8. Commit Changes
model_index += 1
current_config.num_ctx = context_steps[0]
else:
best_configs[track][current_model] = last_good_config
model_index += 1
```
### 9. Commit to Repo
```bash ```bash
cd /opt/data/infra cd /opt/data/infra
git add assets/ai-optimizer/ git add assets/ai-optimizer/
@@ -139,15 +147,6 @@ git commit -m "ai-optimizer: tested ${model} at ${num_ctx} ctx - ${status}"
git push git push
``` ```
### 10. Matrix Notification (if available)
```python
import os
if os.getenv("MATRIX_HOME_SERVER") and os.getenv("MATRIX_ACCESS_TOKEN"):
# Send notification
pass
# Else: silent
```
--- ---
## State File Structure ## State File Structure
@@ -155,7 +154,7 @@ if os.getenv("MATRIX_HOME_SERVER") and os.getenv("MATRIX_ACCESS_TOKEN"):
```json ```json
{ {
"track": "gpu", "track": "gpu",
"current_model": "devstral-small-2:24b", "current_model": "deepseek-coder-v2:16b",
"model_index": 0, "model_index": 0,
"phase": "context_scaling", "phase": "context_scaling",
"backend": "ollama", "backend": "ollama",
@@ -169,10 +168,10 @@ if os.getenv("MATRIX_HOME_SERVER") and os.getenv("MATRIX_ACCESS_TOKEN"):
"ram": {} "ram": {}
}, },
"completed_models": [], "completed_models": [],
"gpu_queue": ["devstral-small-2:24b", "qwen2.5-coder:32b", "codellama:34b-instruct"], "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"], "ram_queue": ["qwen2.5:72b", "nemotron-3-nano:30b", "mixtral:8x7b-instruct"],
"context_steps": [32768, 65536, 98304, 131072, 163840, 200704, 262144, 327680], "context_steps": [32768, 65536, 98304, 131072, 163840, 200704, 262144, 327680],
"last_updated": "2026-04-28T17:00:00Z" "last_updated": "2026-04-30T00:00:00Z"
} }
``` ```
@@ -182,33 +181,29 @@ if os.getenv("MATRIX_HOME_SERVER") and os.getenv("MATRIX_ACCESS_TOKEN"):
```csv ```csv
timestamp,track,model,backend,phase,num_ctx,num_gpu,flash_attn,tokens_per_sec,vram_gb,ram_gb,status,is_best timestamp,track,model,backend,phase,num_ctx,num_gpu,flash_attn,tokens_per_sec,vram_gb,ram_gb,status,is_best
2026-04-28T17:00:00Z,gpu,devstral-small-2:24b,ollama,context_scaling,65536,99,true,15.2,52.1,18.4,success,false 2026-04-30T00:00:00Z,gpu,deepseek-coder-v2:16b,ollama,context_scaling,65536,99,true,15.2,52.1,18.4,success,false
``` ```
--- ---
## Stop Conditions ## Skill Usage
1. All models in both queues have `best_configs` recorded Once PR #1 (ai-worker-restricted-access) is merged:
2. Manual intervention needed (error in state.json `error` field)
3. No progress for 3 consecutive runs
--- ```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>
## Error Handling # Or run the skill directly
ollama-benchmark --model deepseek-coder-v2:16b --track gpu
If any step fails: ```
1. Log error: `"error": {"message": "...", "timestamp": "..."}`
2. Do NOT increment model_index (retry next run)
3. Commit state with error field
4. Exit gracefully
--- ---
## Notes ## Notes
- **No num_parallel**: Removed to avoid limiting other settings - **Manual execution only** - No cron job, run when needed
- **Two tracks**: Complete GPU track first, then RAM track - **Two tracks**: Complete GPU track first (coding models), then RAM track
- **Backend**: Start with ollama, llama.cpp optional - **Backend**: ollama (llama.cpp optional for advanced users)
- **Host access**: Use docker exec or SSH for rocm-smi - **Host access**: Use docker exec or SSH for rocm-smi
- **Ask before deploy**: Show diff before `nh os switch` - **Commit results**: Push best configs to repo for reference

6
assets/ai-optimizer/state.json
View File

@@ -1,6 +1,6 @@
{ {
"track": "gpu", "track": "gpu",
"current_model": "devstral-small-2:24b", "current_model": "deepseek-coder-v2:16b",
"model_index": 0, "model_index": 0,
"phase": "context_scaling", "phase": "context_scaling",
"backend": "ollama", "backend": "ollama",
@@ -14,8 +14,8 @@
"ram": {} "ram": {}
}, },
"completed_models": [], "completed_models": [],
"gpu_queue": ["devstral-small-2:24b", "qwen2.5-coder:32b", "codellama:34b-instruct"], "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"], "ram_queue": ["qwen2.5:72b", "nemotron-3-nano:30b", "mixtral:8x7b-instruct"],
"context_steps": [32768, 65536, 98304, 131072, 163840, 200704, 262144, 327680], "context_steps": [32768, 65536, 98304, 131072, 163840, 200704, 262144, 327680],
"last_updated": "2026-04-28T17:00:00Z" "last_updated": "2026-04-30T00:00:00Z"
} }