我们在RTX 3090上运行Qwen3.5-27B，获得了207 tok/s的性能

Speculative inference for heterogeneous machines and consumer GPUs.
Custom kernels, speculative prefill and decoding, tuned for each model and hardware target.

Inference Engine Optimizations

Supported Models and Drafters

Each badge opens the exact weights used by the measured setup. Drafter badges open the published quant, or the source checkpoint when conversion is required.

Model	🤗 Weights	Optimization (blog)	Phase	Result
Qwen 3.5 0.8B			Prefill + decode	1.9× prefill; 1.55× decode
Qwen 3.8 27B on R9700			Decode	6.4× vs Lucebox AR; 3.8× vs llama.cpp with the same drafter
Laguna XS 2.1 33B			Prefill	6.1×, 411 s to 67.3 s at 256K
Laguna XS 2.1 33B			Decode	1.7× at 256K
Gemma 4 26B-A4B			Decode	1.31×
Gemma 4 31B IT			Decode	3.2×
DeepSeek V4 Flash on Strix Halo			Decode	42 tok/s at 8K and 39 tok/s on code and math with the plain launch (PR #729)
DeepSeek V4.1 Flash R9700 + Strix Halo + SSD, --profile ds41-lucebox			Decode	25.4-25.8 tok/s on code, 16.8-17.5 tok/s after a 4K prompt, at the profile's 128K context (guide)
Ling 3.0 Flash 124B-A5.1B on DGX Spark		Autoregressive	Decode	34.6 tok/s median AR
Qwen 3.8 27B Vision on R9700			Image questions	2.4× vs llama.cpp on DGX Spark at 8 users (10.5 s vs 25.4 s); 2.1× for one user
DeepSeek V4 Flash Vision encoder on R9700, 4 users at once			Image questions	1.5× sooner first token with 16 images when the R9700 encodes

Tested Machines (GPU/APU)

The engine is not tied to one reference card. NVIDIA architectures are selected by CMake; HIP builds should target the device's exact gfx architecture.

	Architecture	Hardware	Runtime	Details
	RDNA4 gfx1201	Radeon AI PRO R9700	ROCm 7.2	Qwen 3.8 R9700 quick start
	RDNA3.5 gfx1151	Ryzen AI MAX+ 395 / Strix Halo	ROCm 7.2	DeepSeek V4 Strix profile
	RDNA3 gfx1100	Radeon RX 7900 XT / XTX	ROCm 6+	DeepSeek V4 dual AMD profile
	Ampere sm_86	RTX 3090	CUDA 12+	Qwen 3.8 NVLink result and Megakernel results
	Blackwell sm_120	RTX 5090	CUDA 12.8+	Qwen 3.8 single-GPU result
	Blackwell sm_121	DGX Spark / GB10	CUDA 12.9	Qwen 3.5 NVFP4 results
	Ada sm_89	RTX 4090	CUDA 12+	Linux and WSL2 community runs
	Turing sm_75	RTX 2080 Ti	CUDA 12.0	DFlash results
	Volta sm_70, Pascal sm_61	V100, P40	CUDA 12.0	CUDA quick start
Not pictured	Blackwell sm_110	Jetson AGX Thor	CUDA 13.0	Thor quick start

Single-device results

Heterogeneous and parallel results

These runs use different prompts, quantizations, and inference policies. They show which configurations work; they are not a cross-hardware ranking.

Recommended Setups

See Recommended server setups for the model and hardware matrix, including single-GPU and mixed-GPU profiles.

The DS4 guide also documents the Strix long-context sparse-verifier profile and Qwen3-0.6B PFlash integration. PFlash is lossy prompt compression; keep it off for exact-retrieval and matched true-context benchmarks.

Client Harnesses

harness/ runs Lucebox through popular coding clients and checks server compatibility.

Set the server binary and model paths, then run a launcher:

LUCE_SERVER_BIN=server/build/luce_server \
LUCE_TARGET=server/models/Qwen3.8-27B-UD-IQ4_XS.gguf \
LUCE_DRAFT=server/models/draft/qwen38-dflash2-q8_0.gguf \
MAX_CTX=32768 \
harness/clients/run_codex.sh

See the harness guide for setup, no-draft targets, and benchmarks.

Quick Start With Docker

Prebuilt images on GHCR track main. Mount the weights and serve the OpenAI-compatible API on :8000.

GPU Image tag NVIDIA (CUDA 12+) :cuda12 AMD (ROCm 6+) :rocm Put the target in server/models/ and its matching drafter in server/models/draft/.

Run the image for your GPU:

# NVIDIA
docker run --rm --gpus all -p 8000:8080 \
  -v "$PWD/server/models:/opt/lucebox-hub/server/models" \
  ghcr.io/luce-org/lucebox-hub:cuda12

# AMD
docker run --rm --device /dev/kfd --device /dev/dri \
  --group-add video --group-add render --security-opt seccomp=unconfined \
  -p 8000:8080 -v "$PWD/server/models:/opt/lucebox-hub/server/models" \
  ghcr.io/luce-org/lucebox-hub:rocm

The container picks the GPU the model fits on (a discrete card before an integrated one, else the largest) and sizes the context from that GPU's memory. serve is the default command, so luce_server flags can follow the image name directly (or serve); they replace the values the container would pass. --target-device, --max-ctx and --profile all work:

# Show the GPUs, the model, and the device auto placement would use
docker run --rm <gpu flags> -v ... ghcr.io/luce-org/lucebox-hub:rocm devices

# DeepSeek V4 on Strix Halo with its qualified profile (DSpark drafter in models/draft/)
docker run --rm <gpu flags> -p 8000:8080 -v ... ghcr.io/luce-org/lucebox-hub:rocm --profile ds4-strix

# Pin a device
docker run --rm <gpu flags> -p 8000:8080 -v ... ghcr.io/luce-org/lucebox-hub:rocm --target-device hip:1

Environment variables such as LUCE_TARGET, LUCE_TARGET_DEVICE, LUCE_MAX_CTX and LUCE_ARGS cover the same settings for compose files; see the header of server/scripts/entrypoint.sh.

Run the Server

This quick start runs the R9700 profile above. The complete flag reference is in the server guide.

# build (ROCm 7.2+, RDNA4)
git clone --recurse-submodules https://github.com/Luce-Org/lucebox.git
cd lucebox
cmake -S server -B server/build-hip -G Ninja \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_HIP_COMPILER=/opt/rocm/lib/llvm/bin/clang++ \
  -DLUCE_GPU_BACKEND=hip \
  -DLUCE_HIP_ARCHITECTURES=gfx1201 \
  -DGGML_HIP_MMQ_MFMA=ON \
  -DGGML_HIP_NO_VMM=ON
cmake --build server/build-hip --target luce_server -j"$(nproc)"

# target and DFlash2 drafter
mkdir -p models
huggingface-cli download unsloth/Qwen3.8-27B-GGUF \
  Qwen3.8-27B-UD-IQ4_XS.gguf --local-dir models
huggingface-cli download incoai/Qwen3.8-27B-DFlash2 --local-dir models/dflash2
python server/scripts/convert_dflash_to_gguf.py \
  models/dflash2/model.safetensors models/qwen38-dflash2-f16.gguf
python server/scripts/quantize_dflash_draft.py \
  models/qwen38-dflash2-f16.gguf models/qwen38-dflash2-q8_0.gguf --scheme q8_0

# launch the measured profile
./server/build-hip/luce_server models/Qwen3.8-27B-UD-IQ4_XS.gguf \
  --draft models/qwen38-dflash2-q8_0.gguf \
  --draft-block-size 16 --max-ctx 131072 \
  --cache-type-k q8_0 --cache-type-v q8_0 \
  --port 8216

curl -s http://127.0.0.1:8216/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"messages":[{"role":"user","content":"Write a Python LRU cache."}],
       "max_tokens":256,"temperature":0}'

To serve up to N concurrent requests, use this launch command with the same Qwen target and DFlash2 drafter. Set N to the desired concurrency (5 below). Qwen automatically sizes the shared KV pool from available GPU memory.

N=5
./server/build-hip/luce_server models/Qwen3.8-27B-UD-IQ4_XS.gguf \
  --draft models/qwen38-dflash2-q8_0.gguf \
  --draft-block-size 16 --max-ctx 16384 \
  --paged-attention --max-concurrency "$N" \
  --cache-type-k q8_0 --cache-type-v q8_0 \
  --port 8216

See Continuous batching in Lucebox for Qwen and DeepSeek V4 Flash results, latency measurements, and launch settings.

Documentation

Benchmarks stay with each implementation: DFlash, PFlash, Spark, KVFlash, and Megakernel.

Tutorials

Video tutorials for each optimization and the harness setup.

The Lucebox Machine

Local AI should be the default, not a privilege. Private data, no per-token bill, no vendor lock-in. Lucebox pairs the R9700 with Strix Halo and ships this open engine ready to run.

See the hardware and current benchmarks at lucebox.com.

Request for Contributions

We welcome focused contributions to CUDA and HIP kernels, speculative inference, support for more consumer GPUs and APUs, performance benchmarks, and client harnesses.

Citation

@software{lucebox_2026,
  title  = {Lucebox: Speculative inference for heterogeneous consumer hardware},
  author = {Lucebox},
  url    = {https://github.com/Luce-Org/lucebox},
  year   = {2026}
}

Community

• Discord: discord.gg/yHfswqZmJQ
• Website: lucebox.com
• Issues: github.com/Luce-Org/lucebox/issues
• Blog: lucebox.com/blog

_{Apache 2.0 · Lucebox.com}