ROCmFPX-BUILDER/README.md

ROCmFPX Automated Builds

---

We provide automated nightly and on-demand builds of ROCmFPX (ciru-ai/ROCmFPX) with AMD ROCm™ 7 acceleration based on TheRock.

ROCmFPX is Ciru's inference framework featuring specialized low-bit quantization formats (ROCmFP2, ROCmFP3, ROCmFP4, ROCmFP6, ROCmFP7 DualView, ROCmFP8), specialized HIP kernels, PromptForge serving runtimes, and fast inference on AMD hardware.

Important

All builds include ROCm™ 7 built-in — complete runtime libraries and BLAS kernels are bundled directly into the archives. No separate ROCm™ driver/SDK installation is required on Windows or Linux!

---

🎯 Supported Devices

This build pipeline targets all major AMD GPU architectures:

• gfx1151 (STX Halo APU) — Ryzen AI MAX+ Pro 395, Radeon 8060S
• gfx1150 (STX Point APU) — Ryzen AI 300 Series
• gfx120X (RDNA4 GPUs) — AMD Radeon RX 9070 XT / 9070 / 9060 XT / 9060
• gfx110X (RDNA3 GPUs) — AMD Radeon dGPUs: PRO W7900/W7800/W7700/W7600, RX 7900 XTX/XT/GRE, RX 7800 XT, RX 7700 XT, RX 7600 XT; iGPUs: Radeon 780M/760M/740M
• gfx103X (RDNA2 GPUs) — AMD Radeon dGPUs: RX 6950 XT, 6900 XT, 6800 XT/6800, RX 6700 XT, RX 6600 XT, RX 6500 XT
• gfx90a (CDNA2 GPUs) — AMD Instinct MI210 / MI250
• gfx908 (CDNA1 GPUs) — AMD Instinct MI100

---

🚀 Automated Builds

Our automated GitHub Actions workflow creates nightly builds for:

• Windows and Ubuntu Linux
• 7 GPU Target Families: gfx1151, gfx1150, gfx120X, gfx110X, gfx103X, gfx90a, gfx908
• ROCm™ 7 Built-in: Full runtime libraries and BLAS support

GPU Target	Target Architecture	Ubuntu (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 despite free VRAM, add ttm.pages_limit=12582912 (48 GB) to your kernel command line (e.g. GRUB), run sudo update-grub, then reboot.

---

⚡ ROCmFPX Format Family

ROCmFPX provides native support for ultra-low-bit formats:

Format	Block BPW	Description
Q2_0_ROCMFPX	2.50	S40 2.5-bpw format with frozen codebook and HIP MMQ/MMVQ paths
Q3_0_ROCMFPX	3.50	3.5-bpw format with scale search and packed execution paths
Q4_0_ROCMFP4 / FAST	4.50 / 4.25	Original ROCmFP4 formats with HIP regression coverage
Q6_0_ROCMFPX	6.50	Strix quality recipes and tuned GPU execution
Q7_0_ROCMFPX	7.50	Signed Q7 format and stored representation for DualView
Q8_0_ROCMFPX	8.25	Reference and cross-backend precision baseline

Quantize any model using llama-quantize:

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

---

🧪 Quick Smoketest

1. Download the archive for your OS and GPU target from Releases.
2. Extract the zip file to your preferred folder.
3. Run llama-server or llama-cli:

# General GGUF Model
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 Structure

• .github/workflows/build-rocmfpx.yml: Matrix CI/CD pipeline for building, bundling, and releasing Windows & Ubuntu binaries.
• .github/workflows/test-rocmfpx.yml: CI/CD testing workflow for smoke testing release artifacts.
• .github/actions/test-rocmfpx-build/: Composite action for executing automated validation tests.
• utils/gather_required_libs.py: Dependency resolution utility to discover and copy missing shared libraries.
• docs/manual_instructions.md: Step-by-step instructions for building locally on Windows and Ubuntu.

---

📋 Manual Build Instructions

For building locally from source, see docs/manual_instructions.md.

---

📄 License

This builder repository is licensed under the MIT License. ROCmFPX is developed by Ciru (ciru-ai/ROCmFPX) under the MIT License.