pomazanbohdan/memory-mcp-1file

This release unlocks the full potential of your CPU by introducing hardware-accelerated machine learning calculations and mitigating thread exhaustion.

🧠 CPU Hardware Acceleration (AVX2, MKL, Accelerate)

• Intel MKL Support: candle now compiles with the mkl feature on Linux and Windows, connecting to the Intel Math Kernel Library. This yields a massive 3x-5x speedup for matrix multiplication during embedding generation.
• Apple Accelerate: MacOS builds now compile with the accelerate feature to utilize the blazing fast Apple Accelerate framework on Intel and Apple Silicon chips.
• AVX2 & FMA Instructions: CI/CD now builds the binaries targeting the x86-64-v3 architecture (covering all Intel/AMD CPUs since ~2014). This enables LLVM to emit AVX2 vector instructions, doubling performance on generic loops.

🧵 Thread Over-Subscription Fix

When deploying this server across an async worker pool (like tokio), handling concurrent inference tasks would cause Rayon and MKL to spawn hundreds of threads (N_cores * N_requests), crushing performance with OS context switching (thread thrashing).

• The server now aggressively limits the underlying mathematical workers by pinning RAYON_NUM_THREADS, MKL_NUM_THREADS, and OMP_NUM_THREADS to 2 during startup.
• You can still override these using your own Environment Variables, but the default provides the perfect sweet spot for multi-tenant, high-throughput environments.