exo Adds Native MTP Support for Qwen3.6

exo, a distributed inference cluster tool, has added native Multi-Token Prediction (MTP) support for Qwen3.6-style MLX checkpoints.

Key performance results and features are as follows:

• 27B model performance improvement: With MTP enabled (K=2, K=3), inference speed was about 2x (1.96x~1.97x) faster compared to the baseline.
• Accuracy verification: The target model verifies candidate tokens generated by the MTP head, guaranteeing identical results to the existing method under greedy decoding.
• Optimal settings per model: While the 27B model benefited from higher K values, the 35B-A3B (MoE) model was most efficient at K=1 due to verification and cache costs.

Currently, the native MTP feature is only supported in single-node environments, and automatically falls back to the existing path when the model is distributed across multiple devices.