dflash2

Up to 4× faster LLM decoding on Apple Silicon, lossless. Native MLX port of DeepSeek's DSpark & z-lab's DFlash speculative decoding — Gemma-4, Qwen3.8, Muse-Glimmer, Nemotron, LFM2.5, Ornith-1.0, ternary Bonsai-27B.

Optimized SGLang runtime for Qwen3.8-27B FP8 with DFlash2 and Qwen3.8 Flash-Next NVFP4 with FR-Spec on one NVIDIA RTX PRO 6000 Blackwell 96 GB GPU (SM120): 524K context, HiCache and NIXL.

Serving Qwen3.8-27B-FP8 on a single DGX Spark (GB10): 7.88 to 58.5 tok/s single-stream from decode strategy alone, weights untouched. Speculative decoding and prefix caching benchmarked, plus DFlash 2 — the only Qwen3.8-27B build that can serve it under vLLM.

Pinned stock Qwen3.8-27B NVFP4 + DFlash2 recipe and measured evidence for one NVIDIA DGX Spark

Qwen3.8-27B EXL3 (4.00 bpw) + DFlash2 speculative decoding for ExLlamaV3, validated at 262k context on a 24 GB RTX 3090

Qwen3.8-27B on one DGX Spark GB10: DFlash 2 for 65 tok/s C1 and 120 C4; EAGLE 3/1/4 for ~191 C8, ~330 C16 and ~535 C32; native 262K, pinned SGLang.

A practical recipe for serving GLM-5.3-Flash (320B MoE) on a single 128GB NVIDIA DGX Spark with DFlash2 speculative decoding.

Evidence-first lab for speculative decoding, DFlash/DFlash2, block diffusion, exactness, acceptance, and inference performance

Qwen 3.8 DFlash2: 27B AI Running at 236 TOK/S Locally! - High-speed local prose and code autocomplete powered by speculative decoding with lightweight draft models.

Native macOS control center and local AI agent gateway for Qwen3.8 on Apple Silicon — MLX, DFlash2, OpenAI Responses, Anthropic Messages, Claude Code, Codex, OpenCode and Grok Build.

Reproducible Qwen3.8-27B serving and native rank-1 LoRA experiments on NVIDIA DGX Spark

Qwen3.8-27B Huihui abliterated NVFP4 on one DGX Spark: 12.31 to 57.95 tok/s with SGLang + DFlash2.

Validated Qwen3.8-27B agent-serving profile for one 48 GiB RTX 4090: DFlash2, KVarN, prefix caching, long-context fairness, and reproducible benchmarks.

Reproducible 262K GLM-5.3 Flash Uncensored EXL3+DFlash2 deployment for 2x RTX PRO 6000 Blackwell

Qwen3.8-27B + DFlash2 block-diffusion speculative decoding on Apple Silicon: 809 tok/s, prefix-cached agentic serving at 262k context

Dual DGX Spark TP=2 serving for GLM-5.3 Flash with an NVFP4 target and MXFP8 DFlash2 draft

Qwen3.8-27B ARA abliterated NVFP4 + DFlash2 on one DGX Spark, with pinned evidence and hardened serving

Standalone eugr/spark-vllm-docker recipes for Qwen3.8-27B NVFP4 and official FP8 on one DGX Spark with dflash2

Educational implementation and visualization of autoregressive, DFlash-inspired, and DFlash2-inspired LLM decoding with a tiny Transformer benchmark.

Evidence-backed GLM-5.3-Flash NVFP4 TP4 deployment validation on 4x NVIDIA GB10.