lucebox-BUILDER (CUDA 12 & AMD ROCm™ 7)

High-performance automated nightly and on-demand builds of lucebox (DFlash inference server) with built-in CUDA 12.8 & AMD ROCm™ 7 runtime libraries.

---

⚡ Project Overview

lucebox (Luce-Org/lucebox) is a high-performance inference engine for DeepSeek-V4, DeepSeek-V3, and DFlash architectures, featuring flash-prefill, speculative decoding, continuous batching, and low-latency KV cache management.

This repository (Heretek-AI/lucebox-BUILDER) delivers turnkey, relocatable binary releases with embedded runtime libraries and $ORIGIN RPATHs — zero host CUDA/ROCm installations required.

flowchart LR
    subgraph Toolchains ["Multi-Backend Toolchains"]
        C1["NVIDIA CUDA 12.8<br/>Fat Binary (sm_75 - sm_120)"]
        C2["AMD ROCm 7 Nightlies (TheRock)<br/>(gfx1151, gfx1100, gfx1201)"]
    end

    subgraph Builder ["lucebox-BUILDER Pipeline"]
        B1["Automated Nightly CI (13:00 UTC)"]
        B2["$ORIGIN RPATH Dynamic Bundler"]
        B3["Non-Blocking Hardware Smoke Tests"]
    end

    subgraph Releases ["Turnkey Binary Releases"]
        R1["lucebox-b####-linux-cuda12-x64.zip"]
        R2["lucebox-b####-linux-rocm-gfx1151-x64.zip"]
        R3["lucebox-b####-linux-rocm-gfx1100-x64.zip"]
        R4["lucebox-b####-linux-rocm-gfx1201-x64.zip"]
    end

    Toolchains --> Builder --> Releases

Loading

---

🎯 Supported Hardware Matrix

Every release ships Linux x86_64 zip packages. ROCm archives bundle all essential shared libraries (libamdhip64, librocblas, libhipblas, libhipblaslt, .kpack, Tensile blobs) with $ORIGIN RPATH:

Asset Name	Backend	Target Architecture & Devices
lucebox-${TAG}-linux-cuda12-x64.zip	CUDA 12.8 Fatbin	sm_75, sm_80, sm_86, sm_89, sm_90, sm_120 (RTX 2080 Ti → RTX 3090 / 4090 / 5090, A100, H100)
lucebox-${TAG}-linux-rocm-gfx1151-x64.zip	AMD HIP / ROCm 7	AMD Strix Halo APU (Ryzen AI MAX+ Pro 395 / Radeon 8060S / 128GB Unified Memory)
lucebox-${TAG}-linux-rocm-gfx1100-x64.zip	AMD HIP / ROCm 7	AMD RDNA3 Discrete GPUs (Radeon RX 7900 XTX / 7900 XT / 7800 XT)
lucebox-${TAG}-linux-rocm-gfx1201-x64.zip	AMD HIP / ROCm 7	AMD RDNA4 Discrete GPUs (Radeon AI PRO R9700, RX 9070 XT / 9070)

---

📦 Quick Start

# 1. Download and extract target archive from GitHub Releases
unzip lucebox-b1000-linux-rocm-gfx1151-x64.zip -d lucebox
cd lucebox/bin

# 2. Launch DFlash inference server
./dflash_server /path/to/DeepSeek-V4-Flash.gguf --port 8080 --host 0.0.0.0

# 3. Speculative decoding with draft model
./dflash_server main_model.gguf --draft draft_model.gguf --port 8080

---

🚀 CI Workflows

.github/workflows/build-lucebox.yml:

• Nightly cron at 13:00 UTC, plus manual workflow_dispatch (configurable CUDA arches, gfx targets, ROCm version, and rocWMMA Phase-2 kernels).
• TheRock ROCm Toolchain: Downloads nightly multi-arch tarballs from AMD's official distribution mirrors.
• CUDA 12.8 Toolchain: Builds fat binaries covering compute capabilities from Turing (sm_75) through Blackwell (sm_120).
• Self-Hosted Smoke Tests: Non-blocking verification on live GPU hardware.

---

📄 License & Attribution

• lucebox: Developed by Luce Org (Luce-Org/lucebox) under the MIT License.
• lucebox-BUILDER Pipeline: Developed by Heretek-AI under the MIT License.