We got 207 tok/s with Qwen3.5-27B on an RTX 3090

Open LLM inference, rewritten by hand for one specific chip at a time.

Kernels, speculative decoding, and quantization, tailored per target.

We don't wait for better silicon. We rewrite the software.

Two projects today, more coming. Each one is a self-contained release with its own benchmarks and paper-style writeup.

The first megakernel for hybrid DeltaNet/Attention LLMs. All 24 layers of Qwen 3.5-0.8B in a single CUDA dispatch, 1.87 tok/J on a 2020 GPU, matching Apple's latest silicon at 2× the throughput.

# 1. clone + enter
git clone https://github.com/Luce-Org/lucebox-hub && cd lucebox-hub/megakernel
# 2. install (Python 3.10+, CUDA 12+, PyTorch 2.0+). Weights stream from HF on first run.
pip install -e .
# 3. run the benchmark (prefill pp520 + decode tg128 vs llama.cpp BF16 + PyTorch HF)
python final_bench.py

| Method | Prefill pp520 | Decode tg128 | tok/J | |---|---|---|---| | Megakernel @220W | 37,800 | 413 | 1.87 | | llama.cpp BF16 @350W | 11,247 | 267 | 0.76 | | PyTorch HF | 7,578 | 108 | n/a |

What makes it work: 82 blocks, 512 threads, one persistent kernel. No CPU round-trips between layers. Weights streamed straight from HuggingFace. Cooperative grid sync instead of ~100 kernel launches per token. Power ceiling hit before compute ceiling, so DVFS converts tight execution straight into saved watts.

First GGUF port of DFlash speculative decoding. Qwen3.5-27B on a single RTX 3090, Q4_K_M target + BF16 draft, DDTree budget=22.

Up to 207 tok/s in the demo (207.6 tok/s DFlash vs 38.0 tok/s AR, 5.46×)
129.5 tok/s mean on the HumanEval 10-prompt bench
3.43× faster than autoregressive (+15% over chain speculative decoding)
2.8× faster than SGLang AWQ on the same hardware
128K context in 24 GB (134.78 tok/s at ctx=131072)

# 1. clone with submodules (pulls the pinned Luce-Org/llama.cpp@luce-dflash fork)
git clone --recurse-submodules https://github.com/Luce-Org/lucebox-hub && cd lucebox-hub/dflash
# 2. build the C++/CUDA decoder (~3 min on sm_86, CUDA 12+, CMake 3.18+)
cmake -B build -S . -DCMAKE_CUDA_ARCHITECTURES=86 -DCMAKE_BUILD_TYPE=Release
cmake --build build --target test_dflash -j
# 3. fetch weights: ~16 GB Q4_K_M target + 3.46 GB bf16 draft
huggingface-cli download unsloth/Qwen3.5-27B-GGUF Qwen3.5-27B-Q4_K_M.gguf --local-dir models/
huggingface-cli download z-lab/Qwen3.5-27B-DFlash model.safetensors --local-dir models/draft/
# 4a. one-shot streaming generate
python3 scripts/run.py --prompt "def fibonacci(n):"
# 4b. or reproduce the paper-style bench (HumanEval + GSM8K + Math500, ~15 min)
python3 scripts/bench_llm.py

| Benchmark | AR (tok/s) | DFlash+DDTree (tok/s) | Speedup | |---|---|---|---| | HumanEval | 37.8 | 129.5 | 3.43× | | Math500 | 37.7 | 110.5 | 2.93× | | GSM8K | 37.7 | 96.2 | 2.55× |

The constraint that shaped the project. AWQ INT4 of Qwen3.5-27B plus the BF16 draft doesn't leave room for the DDTree verify state on a 24 GB card. Q4_K_M GGUF (~16 GB target) is the largest format that fits target + 3.46 GB draft + budget=22 tree state + KV cache in 24 GB on the RTX 3090. Picking it forced a new port on top of ggml, since no public DFlash runtime supports a GGUF target.

What we built vs what we didn't. The algorithms are not ours:

• DFlash (z-lab, 2026): block-diffusion draft conditioned on target hidden states.
• DDTree (Ringel et al., 2026): tree-structured verify that beats chain verify at the same compute budget.

What we ported and tuned:

• C++/CUDA decode engine on top of ggml (no libllama, no Python runtime, Q4_K_M target path).
• Three custom CUDA kernels for tree-aware SSM state rollback.
• DDTree budget swept for RTX 3090 + Q4_K_M target: budget=22 is the sweet spot.
• Q4_0 KV cache + sliding target_feat ring to fit 128K context in 24 GB.

Local AI should be a default, not a privilege. General-purpose frameworks dominated the last decade because hand-tuning kernels per chip was too expensive. AI-assisted development flips that calculus. Lucebox is where we publish these rewrites, one chip and one model family at a time.