https://ramdi.fr/tags/speculative-decoding/Recent content in Speculative-Decoding on Noureddine RAMDIHugoenSat, 23 May 2026 20:41:27 +0000https://ramdi.fr/github-stars/dflash-mlx-speculative-decoding-on-apple-silicon-with-metal-and-mlx/Mon, 04 May 2026 10:23:02 +0000https://ramdi.fr/github-stars/dflash-mlx-speculative-decoding-on-apple-silicon-with-metal-and-mlx/dflash-mlx implements exact speculative decoding for language models on Apple Silicon using Metal and MLX, reducing forward passes with a block-diffusion draft model and per-layer KV cache rollback.https://ramdi.fr/post/ai-llm/local-llm-tokens-per-second-benchmark/Tue, 28 Apr 2026 00:00:00 +0000https://ramdi.fr/post/ai-llm/local-llm-tokens-per-second-benchmark/Part 2 of 4: a benchmark journal across nixpkgs llama.cpp, upstream master, and ik_llama.cpp on Qwen3.6-27B. Six hours, four backends, all converging at 66 tok/s — and the physical reason why.https://ramdi.fr/post/ai-llm/local-llm-speculative-decoding-hybrid-ssm/Tue, 28 Apr 2026 00:00:00 +0000https://ramdi.fr/post/ai-llm/local-llm-speculative-decoding-hybrid-ssm/Part 3 of 4: a deep-dive into why speculative decoding silently breaks (or runs anti-economically) on hybrid attention+SSM architectures like Qwen3.6, Mamba-2, and RWKV — and what would need to change upstream to fix it.