magnitudedev/magnitude @magnitudedev/cli@0.2.10

on GitHub

0.2.10

Patch Changes

• 15ff695 Thanks @anerli! - Fix memory planning on unified-memory and integrated GPUs.

  • On NVIDIA GB10 and Apple Silicon, the one-time optimization counts the memory the loaded model already uses instead of overcommitting it (#199, #206).
  • Integrated GPUs such as AMD Strix Halo use their shared memory as well as the dedicated carve-out, instead of stopping when the carve-out is full (#207).
  • The optimization retries an allocation the GPU refuses instead of failing.
  • The app lists each GPU once and shows the backend Magnitude uses for it.

• 7f02906 Thanks @anerli! - Support models with IQ2_XXS weights on every backend, and fix models with IQ2_XS weights failing to load on a Mac (#195).

• 7ec9327 Thanks @thrgreenwald! - Fix model loads failing on Linux machines and virtual machines (#196).

  • Memory held by the file cache counts as available, so loads no longer fail for lack of memory that the system could free.
  • A real memory shortage at startup is reported as one that can be retried, instead of failing the load partway through.
  • A CPU without the instructions the engine needs is reported plainly before loading starts.

• 151bf9a Thanks @anerli! - Fix models with a Q8_0 output head failing to load, including MiniCPM5 1B Q8 and, with drafting off, Qwen3.6 35B-A3B and Qwen3.5 9B Q6. Faster output for large vocabularies can be selected again during the one-time optimization.

• 1f54d5b Thanks @thrgreenwald! - The web app's model status explains that a model loads automatically on the first request, and the Update button explains what an update brings.

• fc5e2e8 Thanks @thrgreenwald! - Fix the RPM package conflicting with other installed applications over shared build IDs (#200).