SDXL Pony LoRA training system with dataset management, AI tagging, and in-browser testing
- Python 63.9%
- HTML 26.6%
- CSS 5.5%
- JavaScript 2.4%
- Dockerfile 1.6%
jbelec
e3ce067768
The subprocess PIPE buffer (64KB) fills up when accelerate launch's nested processes produce output faster than async readline drains it. This causes the child to block on write() and training stalls. Fixed by redirecting stdout directly to the log file (no PIPE). The monitor task now tails the log file every 2s for progress parsing and SSE streaming. This eliminates the pipe bottleneck entirely. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|---|---|---|
| config | ||
| src | ||
| vendor | ||
| .gitignore | ||
| .gitmodules | ||
| docker-compose.yml | ||
| Dockerfile | ||
| requirements.txt | ||