Laya on Apple Silicon: worker, benchmarks and recipe - #30

Open

cacheline999 wants to merge 24 commits into

ThinkFlowLab:mainfrom

cacheline999:laya-apple-silicon

Commits on Sep 30, 2026

Add LAYA_WORKER_WEIGHTS=fp16 for the compiled worker
cacheline999
committed
d3d0817
Compile the encoder for multi-question batches; make compile on/off
cacheline999
committed
e10cf06
Recipe: M5 results, test dependencies, first-request note
cacheline999
committed
b00b285
Say that fp16 weights are for MPS and warn on CPU
cacheline999
committed
3dc662b
Read /health from the live model; correct the Apple Silicon recipe
cacheline999
committed
6311ae8
Run plain fp32 laya after a fallback to the CPU
cacheline999
committed
9ad04e3
Split the GPU optimizations out of worker.py; correct the Laya README
cacheline999
committed
5e260a1
Merge remote-tracking branch 'origin/main' into laya-apple-silicon
cacheline999
committed
01db6e7

Commits on Oct 2, 2026