otf-llm 3.2.0

High-performance hybrid LLM inference engine with Fused OpenAI Triton INT4 GEMM kernels, 98.16% logit parity, Zero-RAM streaming quantizer, and 3-Tier MoE offloading.