Spaces:

---

prithivMLmods

/

Qwen-Image-Edit-2511-LoRAs-Fast

like 1.68k

Running on Zero

The purpose of these changes is to stop LoRA adapters from accumulating in GPU memory forever.

Before:

• Every newly selected LoRA stayed loaded for the lifetime of the app.
• VRAM usage could keep growing as users tried more adapters.
• Because the pipeline is global, concurrent requests could also interfere with adapter switching.

After:

• The app keeps only a small number of recently used LoRAs in memory.
• When the cache is full, it evicts the least recently used adapter before loading a new one.
• A lock ensures adapter load/evict/switch operations don’t race with inference.

So the practical goal is:

• lower long-session VRAM growth
• make the app more stable under repeated adapter switching
• reduce the chance of OOMs and shared-pipeline corruption

You can control the cache size with MAX_CACHED_LORAS.