v0.16.4

Elastic VRAM, live RAM. Everything in v0.16.0 through v0.16.3 still applies.

New

• Elastic VRAM (Setup tab, off by default, NVIDIA only). When another app needs your GPU, LlamaForge moves a loaded model's layers to the CPU instead of unloading it (MoE experts first), keeps the conversation through llama.cpp's slot save and restore, and moves the layers back once the memory is free.
  • A resized model is much slower; the dashboard row says so, and its chip grows it back on click.
  • Keeping conversations needs one router restart (the router only gets a slot-save folder while Elastic is on).
  • New and not yet tested on many machines. Reports welcome.
• Live system RAM on Stats, next to GPU memory, in every look, and on machines without a GPU too. "Used" leaves out cache the OS can hand back, like Task Manager's "In use". Asked for on Reddit.
• A warning before an unknown architecture fails. If a model is newer than the llama.cpp checkout that would load it, the model editor tells you to update the build instead of the load failing with unknown model architecture. Embedding models such as EmbeddingGemma now have a section in the models.ini docs.

Fixes

• /v1/audio/speech answers an unsupported response_format (say mp3) with a 400 in the OpenAI error shape instead of WAV under the wrong name. Reported by u/Natethegreat9999.
• Ember quotes copied from PDFs now match across line-break and soft hyphens; paraphrases still fail. Raised by u/NeuralNomad87.
• Reloading models.ini works on router builds that serve it on a different route, and allowed values: lists that wrap over several lines in --help are read whole. Thanks @Vergillll (#39, #40).

Already on v0.16.x? The panel offers this update under Discover → New this week (installer copies; a git checkout updates with git pull).