openclaw-voice/PYTORCH_MONITORING.md
MCKRUZ 74167edc0d docs: Add PyTorch RTX 5090 monitoring guide
- Weekly check instructions for sm_120 support
- Automated monitoring options (GitHub Watch, RSS, calendar)
- Step-by-step GPU setup when support lands
- Historical timeline estimates (2-3 months typical)
- Optimistic timeline: March 2025
- While-waiting optimization suggestions

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-02-16 19:54:54 -05:00

220 lines
4.8 KiB
Markdown

# PyTorch RTX 5090 Support Monitoring Guide
**Goal:** Get notified when PyTorch adds Blackwell (sm_120) support
---
## Quick Check (Weekly)
### Test PyTorch Nightly
```bash
# From project root
cd "C:\Users\kruz7\OneDrive\Documents\Code Repos\MCKRUZ\openclaw-voice"
source venv/Scripts/activate
# Test nightly build
pip install --upgrade --pre torch --index-url https://download.pytorch.org/whl/nightly/cu124
# Quick test
python -c "import torch; x=torch.rand(10,10,device='cuda'); print('✓ RTX 5090 WORKS!' if x.device.type=='cuda' else '✗ Not yet')"
```
If you see **"✓ RTX 5090 WORKS!"** → GPU support is here! Run `fix_pytorch_cuda.bat`
---
## Automated Monitoring
### Option 1: GitHub Watch (Recommended)
1. Go to: https://github.com/pytorch/pytorch
2. Click **"Watch"** → **"Custom"**
3. Check **"Releases"** only
4. Get email when new PyTorch releases
### Option 2: RSS Feed
Subscribe to PyTorch releases:
```
https://github.com/pytorch/pytorch/releases.atom
```
Use RSS reader (Feedly, Inoreader) or browser extension
### Option 3: Weekly Calendar Reminder
Set recurring calendar event:
- **When:** Every Monday 9am
- **What:** Check PyTorch RTX 5090 support
- **How:** Run quick test above
---
## What to Look For
### In Release Notes
Keywords indicating sm_120 support:
- ✅ "Blackwell" or "sm_120"
- ✅ "RTX 5090" or "50-series"
- ✅ "CUDA capability 12.0"
- ✅ "Hopper+Blackwell" or "H100+B100"
### Example Release Note:
```
PyTorch 2.X.0 Release Notes
- Added support for NVIDIA Blackwell architecture (sm_120)
- RTX 50-series GPUs now fully supported
```
---
## When Support Lands
### 1. Update PyTorch
```bash
cd "C:\Users\kruz7\OneDrive\Documents\Code Repos\MCKRUZ\openclaw-voice"
# Run the fix script
fix_pytorch_cuda.bat
# Or manually:
source venv/Scripts/activate
pip uninstall torch torchaudio torchvision -y
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
```
### 2. Verify GPU Works
```bash
python -c "import torch; print(torch.cuda.get_device_name(0)); x=torch.rand(100,100,device='cuda'); print('GPU OK!')"
```
Expected output:
```
NVIDIA GeForce RTX 5090
GPU OK!
```
### 3. Update Config
Edit `config.yaml`:
```yaml
pipeline:
stt:
device: "cuda" # Was: cpu
model_size: "medium" # Can increase from small
beam_size: 5 # Can increase from 1
tts:
device: "cuda" # Was: cpu
```
### 4. Test Performance
```bash
# Start the bot
python run.py
# In Discord
/join
# Say: "Hey Jarvis, test response time"
/status # Check latency stats
# Expected improvement:
# - STT: ~2s → ~0.35s (6x faster)
# - TTS: ~4s → ~0.9s (4x faster)
# - Total: ~10s → ~4s (near 3.5s target!)
```
### 5. Re-test Kani-TTS-2
```bash
python test_kani_tts.py
# If successful:
# - Compare quality with current Coqui XTTS v2
# - Check if RTF ~0.2 achieved
# - Decide if worth integrating
```
---
## Estimated Timeline
Based on historical GPU support addition:
| GPU Architecture | Release Date | PyTorch Support Added | Time Gap |
|------------------|--------------|----------------------|----------|
| Ampere (RTX 30) | Sep 2020 | Nov 2020 | 2 months |
| Ada Lovelace (RTX 40) | Oct 2022 | Dec 2022 | 2 months |
| Hopper (H100) | Mar 2023 | May 2023 | 2 months |
| **Blackwell (RTX 50)** | **Jan 2025** | **Est: Mar-Apr 2025** | **2-3 months** |
**Conservative estimate:** March 2025 (1 month from now)
**Optimistic estimate:** Late February 2025 (2 weeks)
**Pessimistic estimate:** May 2025 (3 months)
---
## While You Wait
### Optimize Non-GPU Components
Focus on improvements that work on CPU:
1. **Query Routing** (already implemented)
- Haiku for simple queries
- Sonnet for medium
- Opus for complex
2. **TTS Caching** (already implemented)
- Pre-generate common phrases
- Cache by hash
3. **Response Filtering**
- Improve relevance detection
- Reduce unnecessary responses
4. **Streaming Optimization**
- Sentence-level playback
- Parallel processing where possible
### Test Bot Logic
Even with slow performance, you can:
- Test conversation flow
- Debug agent personalities
- Refine prompt engineering
- Test Discord commands
- Verify OpenClaw integration
### Prepare for GPU
- Read KANI_TTS_EVALUATION.md
- Plan integration strategy
- Review current TTS implementation
- Identify optimization opportunities
---
## Contact/Support
**PyTorch Issues:** https://github.com/pytorch/pytorch/issues
**PyTorch Forums:** https://discuss.pytorch.org/
**NVIDIA Developer:** https://forums.developer.nvidia.com/
**Search for:** "RTX 5090 support" or "sm_120" or "Blackwell"
---
## Summary
**Weekly:** Run quick test or check GitHub releases
**When ready:** Run `fix_pytorch_cuda.bat`
**Then:** Update config, test performance, evaluate Kani-TTS-2
**Expected:** March 2025 (1-2 months)
**Bookmark this file** and check weekly until GPU support lands!