v0.81.29
Fixed
- External GPUs (eGPU) run at full speed. The Metal backend forces shared
(system-memory) buffers for external GPUs, so weights stream over Thunderbolt every
op (~0.8 t/s). ToshLLM now forces private, VRAM-resident buffers
(GGML_METAL_SHARED_BUFFERS_DISABLE) automatically when an external GPU is selected,
with a manual "VRAM-resident weights" toggle (shown only when an eGPU is present) for
the default case where macOS picks the eGPU. Likely also clears the
failed to decode prompt batch (res = -3)benchmark error on eGPUs.
Added
- Self-contained server log header. Each server start now logs the app version,
engine, model, detected GPUs (flagging external/eGPU), the GPU selection, the resolved
settings and the exact args/env (API key redacted) — so a single pasted log is enough
to debug a remote setup.
Full Changelog: v0.81.28...v0.81.29