OPTIMIZING MODELS TO BE FAST AT CODEGEN

Summary from the TLDR AI newsletter

Morph LLM optimizes open coding models by training a drafter on coding output, not the internet, for faster speculative decoding, achieving a 3.07x speedup. Autoresearch automates kernel tuning for low-demand GPUs like NVIDIA and AMD, enhancing warp-decode kernels to reach 162 tok/s on affordable hardware. Interconnect over PCIe replaces expensive NVLink using custom kernels, maintaining performance by sharing caches via TCP, cutting time-to-first-token by 84%.