Krea 2 Raw Training Slowdown on RTX 5080: Potential AI Toolkit Bug
NexusFred · reddit · 2026-08-23
A user reported severe performance issues when training Krea 2 Raw LoRA on an RTX 5080 16GB. Despite enabling Low VRAM and gradient checkpointing, speed degraded from 21s/iter to 40s/iter with near-max VRAM usage. In contrast, the same model trains much faster on OneTrainer, suggesting a potential bug related to AI Toolkit issue #990.
More from coding & agent
- Opus 5 guide: Set effort to Medium, define precise goals, and let it run — daniel_mac8 · 2026-08-23
- Claude Code Version 2.1.241 Incoming — ClaudeCodeLog · 2026-08-23
- Dev Rants on Cloud Environments: Wants Remote Laptop Experience — jasonkneen · 2026-08-23
- Tsinghua + Cornell: ACID-Transaction framework for reliable Agent memory — rohanpaul_ai · 2026-08-23
- MCP Connector for Hotel Booking Launches with USDC Payments on Base — modelcontextprotocol · 2026-08-23
- Hybrid Approach: Cloud for Isolation, Local for Sovereignty — sull · 2026-08-23