2026-09-21 · America/Los_Angeles · 社区动态 · #18

Run Jev-style typed decisions locally on Apple Silicon

I built Laya MPS, a PyTorch runtime for running Jev-style typed decisions locally on Apple Silicon GitHub repo: https://github.com/afshinm/laya-mps On an M5 Pro: - Default mode: ~32 ms latency, ~2.1 GB peak process RAM - Lowest-memory mode: ~171 ms latency, ~0.74 GB peak process RAM All modes run the same complete model in FP32. Reading embeddings and a different option to run the model layers from disk for lower RAM usage. Includes a local HTTP API and a demo that doubles as a live latency benchmark. The latency figures above use fixed Pong inputs and exclude HTTP overhead.

热度 45.2 / 100；排名与评分保留该期记录。