ROCmFPX Automated Builds (AMD ROCm™ 7)

High-performance automated nightly and on-demand builds of ROCmFPX with built-in AMD ROCm™ 7 runtime libraries for Windows & Ubuntu.

---

⚡ What is ROCmFPX?

ROCmFPX (ciru-ai/ROCmFPX) is Ciru's specialized high-performance AMD inference stack for low-bit quantization formats, mixed-precision execution, and GPU-accelerated runtime paths:

• ROCmFP2 (2.50 bpw), ROCmFP3 (3.50 bpw), ROCmFP4 / FAST (4.50 / 4.25 bpw), ROCmFP6 (6.50 bpw), ROCmFP7 DualView (7.50 bpw), and ROCmFP8 (8.25 bpw).
• DualView Architecture: Authoritative Q7 storage with zero-copy Q7-decode streaming and exact signed-Q8 prefill shadow for maximum throughput.
• ActiveFPX PromptForge: Prompt-specialized runtime featuring fused gate/up projections, merged QKV/Z projections, and shape-guarded routes on Strix Halo (gfx1151).
• HIP/ROCm Acceleration: Tailored MMQ and MMVQ execution paths tuned for AMD RDNA2, RDNA3, RDNA3.5, RDNA4, and CDNA architectures.

Important

⚡ Ready to Run — ROCm™ 7 Built-in: All binaries include complete ROCm 7 runtime libraries, hipBLAS, rocBLAS, and hipBLASLt kernels. No separate AMD ROCm™ SDK or driver installation is required on Windows or Linux!

---

🎯 Supported AMD GPU Targets

Target Code	GPU Architecture	Target Hardware / Devices
gfx1151	Strix Halo APU (RDNA3.5)	AMD Ryzen AI MAX+ Pro 395, Ryzen AI MAX 390, Radeon 8060S
gfx1150	Strix Point APU (RDNA3.5)	AMD Ryzen AI 9 HX 370, Ryzen AI 9 365, Radeon 890M / 880M
gfx120X	RDNA4 dGPUs	AMD Radeon RX 9070 XT, RX 9070 GRE, RX 9070, RX 9060 XT, RX 9060
gfx110X	RDNA3 dGPUs & iGPUs	Radeon PRO W7900 / W7800 / W7700 / W7600, RX 7900 XTX / XT / GRE, RX 7800 XT, RX 7700 XT, RX 7600 XT, Radeon 780M / 760M
gfx103X	RDNA2 dGPUs	AMD Radeon RX 6950 XT, RX 6900 XT, RX 6800 XT / 6800, RX 6700 XT, RX 6600 XT, RX 6500 XT
gfx90a	CDNA2 Accelerators	AMD Instinct MI210, MI250, MI250X
gfx908	CDNA1 Accelerators	AMD Instinct MI100

---

🚀 Download Release Binaries

Automated GitHub Actions builds are published nightly and on every release:

GPU Target	Architecture Family	Ubuntu Linux (x64)	Windows (x64)
gfx1151	Strix Halo (Ryzen AI MAX+ 395)
gfx1150	Strix Point (Ryzen AI 300)
gfx120X	RDNA4 (RX 9070 XT / 9070)
gfx110X	RDNA3 (RX 7900 XTX / 7800 XT / 780M)
gfx103X	RDNA2 (RX 6800 XT / 6700 XT)
gfx90a	CDNA2 (Instinct MI210 / MI250)
gfx908	CDNA1 (Instinct MI100)

Linux (gfx1150/gfx1151 APU): If you experience OOM errors despite free VRAM, add ttm.pages_limit=12582912 (48 GB) to your kernel command line (e.g. GRUB), run sudo update-grub, and reboot.

---

📦 ROCmFPX Format Catalog

ROCmFPX introduces dedicated GGUF low-bit quantization formats optimized for ROCm MMQ/MMVQ kernels:

Format	Block BPW	Implementation & Features
Q2_0_ROCMFPX	2.50	S40 2.50-bpw format with frozen codebook and HIP MMQ/MMVQ dispatch
Q3_0_ROCMFPX	3.50	3.50-bpw format with scale-search and packed execution paths
Q4_0_ROCMFP4 / FAST	4.50 / 4.25	Original ROCmFP4 formats with regression coverage and serving stability
Q6_0_ROCMFPX	6.50	Strix quality recipes with optimized HIP/ROCm GPU execution
Q7_0_ROCMFPX	7.50	Signed Q7 format and stored representation for DualView
Q8_0_ROCMFPX	8.25	Reference precision and cross-backend baseline

Quantize any model using llama-quantize:

llama-quantize source-model-BF16.gguf output-model-FP7.gguf Q7_0_ROCMFPX

---

🧪 Quick Smoketest & Usage

🍺 Homebrew Installation (Linux)

Install pre-compiled ROCmFPX with embedded ROCm 7 runtime libraries via Homebrew:

# 1. Tap the Heretek-AI repository
brew tap Heretek-AI/tap

# 2. Install ROCmFPX (defaults to AMD Strix Halo gfx1151):
brew install rocmfpx

# Or install for your specific GPU architecture:
brew install rocmfpx --with-gfx1150  # AMD Strix Point (Radeon 890M)
brew install rocmfpx --with-gfx120X  # AMD RDNA4 (RX 9070 XT)
brew install rocmfpx --with-gfx110X  # AMD RDNA3 (RX 7900, Radeon 780M)
brew install rocmfpx --with-gfx103X  # AMD RDNA2 / Steam Deck

# 3. Optional: Run background OpenAI-compatible server daemon:
brew services start rocmfpx

Manual Download & Extraction

1. Download the zip archive matching your GPU from Releases.
2. Extract the archive to any directory.
3. Run llama-server or llama-cli:

# Start llama-server with all GPU layers offloaded
llama-server -m YOUR_MODEL.gguf -ngl 99

Running DualView Models (e.g. Ornith 35B):

export GGML_ROCM_GFX1151_Q7_Q8_VIEW=no-output
export GGML_HIP_ENABLE_UNIFIED_MEMORY=1

./llama-server \
  -m Ornith1.0-35b-CIRU-DUALVIEW-FPX7+Q8-MTP.gguf \
  --host 127.0.0.1 --port 8080 \
  -dev ROCm0 -sm none -ngl 999 -fa on \
  -n 16384 -c 131072 -b 2048 -ub 512 -t 16 -tb 16 \
  -ctk f16 -ctv f16 --parallel 1 --metrics --mmap --no-repack

---

🏗️ Repository Overview

• .github/workflows/build-rocmfpx.yml: Multi-platform matrix build pipeline downloading TheRock ROCm toolchain and publishing release archives.
• .github/workflows/test-rocmfpx.yml: Standalone testing workflow for verifying release artifacts.
• .github/actions/test-rocmfpx-build/: Composite testing action for model inference and GPU offload validation.
• utils/gather_required_libs.py: Dynamic dependency discovery utility.
• docs/manual_instructions.md: Step-by-step local compilation guide.
• AGENTS.md / CLAUDE.md / GEMINI.md: Guidelines for AI coding agents.

---

📄 License

This builder pipeline is licensed under the MIT License.
ROCmFPX is developed by Ciru (ciru-ai/ROCmFPX) under the MIT License. Llama.cpp is developed by Georgi Gerganov and contributors under the MIT License.