Optimizing Diffusion LLMs on Apple Silicon
DunklerErpel · reddit · 2026-07-15
This is an in-depth post detailing experiments with diffusion LLMs (like LLaDA2.1 and Sumi) on Apple Silicon, accompanied by a released GitHub repo named NeoDiffusion.
The content primarily summarizes various optimization experiments, covering baselines to multiple acceleration strategies:
- Quantization schemes: Comparing 4-bit, 6-bit, and other layouts, detailing VRAM usage and flip rates
- Cache optimization: Such as elastic cache and FlashBlock caching; some methods were rejected or require re-benchmarking
- Speculation / Speculative execution: Including self-speculation and draft graphs; certain presets were accepted
- Token freezing / JOT: Improved reasoning throughput but incurred synchronization overhead on some platforms
- Multiple experiments logged metrics like TPS, step count changes, and compute negative/positive
Overall, this is an engineering-focused performance log. The focus isn't on a "model release," but rather on how to run dLLMs faster and more stably on local hardware.
More from Infra
- China’s AI arms race is increasingly defined by chips, data centers, and open models — BenBajarin · 2026-07-22
- Agent search bottlenecks are now about variance, not raw latency — rohanpaul_ai · 2026-07-22
- Gavin Baker argues Nvidia may be one of open source AI’s biggest supporters — GavinSBaker · 2026-07-22
- AI Power Demand Exposes US Energy Gap, Urging Shift from Scarcity to Abundance — bradneuberg · 2026-07-22
- Gavin Baker says Nvidia’s $630B figure would be system revenue, not all Nvidia’s — GavinSBaker · 2026-07-22
- A Firecracker-based platform says it can host 6,000 AI agents on one 256 GB server — maritime_sh · 2026-07-22