GitHub - iggerask/mlx-mtp: MTP speculative decoding for Apple Silicon (MLX) — 1.27x throughput on Qwen3.5-35B-A3B via fused MoE kernels + zero-replay rejection · GitHub

Name		Name	Last commit message	Last commit date
Latest commit History 3 Commits
benchmarks		benchmarks
docs		docs
mlx_fused_moe		mlx_fused_moe
results		results
scripts		scripts
tests		tests
vllm_mlx_mtp		vllm_mlx_mtp
.gitignore		.gitignore
pyproject.toml		pyproject.toml
setup_cython.py		setup_cython.py

About

MTP speculative decoding for Apple Silicon (MLX) — 1.27x throughput on Qwen3.5-35B-A3B via fused MoE kernels + zero-replay rejection

0 stars

0 watching

Report repository

Releases

Packages

Contributors

Languages