Luce-Org's DFlash port with DDTree budget=22 delivers 207.6 tokens/s on Qwen3.5-27B Q4_K_M target plus BF16 draft versus 38.0 tok/s autoregressive baseline, a 5.46× gain, and 129.5 tok/s mean on HumanEval (3.43× speedup). Three custom CUDA kernels (ggml_ssm_conv_tree, ggml_gated_delta_net_tree, ggml_gated_delta_net_tree_persist) enable tree-aware SSM state rollback on ggml backend. Memory engineering with Q4_0 KV cache and sliding target_feat ring fits 128K context plus verify state in 24 GB VRAM (https://github.com/Luce-Org/lucebox-hub).

The release synthesizes z-lab DFlash block-diffusion draft conditioned on target hidden states (2026) and Ringel et al. DDTree structured verification that outperforms chain methods at equal budget (2026), extending Leviathan et al. speculative sampling techniques (https://arxiv.org/abs/2211.17192). Megakernel companion project places all 24 layers of Qwen3.5-0.8B in one CUDA dispatch for 413 decode tok/s at 1.87 tok/J, exceeding llama.cpp BF16 (267 tok/s at 0.76 tok/J) and matching Apple silicon efficiency without new hardware (https://github.com/ggerganov/llama.cpp). Cooperative grid sync removes ~100 kernel launches per token.

Project constraints forced Q4_K_M selection after AWQ INT4 plus draft left insufficient room for DDTree state; this yields 2.8× speedup over SGLang AWQ on identical RTX 3090. Persistent kernel and power-ceiling execution convert latency directly to watt savings, continuing pattern of per-chip hand-tuned kernels rather than framework-level abstractions seen in prior llama.cpp and vLLM releases.