Latest LM Studio update killed MTP performance

In three linesUser reports LM Studio update from 0.4.14 to 0.4.17 degraded MTP (Multi-Token Prediction) performance on RTX 5090. Throughput dropped from ~100 tokens/s with MTP enabled back to ~70 tokens/s after update and CUDA runtime refresh.

Summary generated by Claude — human-verified