ROCmFPX Automated Builds (AMD ROCm™ 7)

High-performance automated nightly and on-demand builds of ROCmFPX (Upstream), Ciru-AI ROCmFPX, and q38rocm with built-in AMD ROCm™ 7 runtime libraries for Windows & Ubuntu.

---

⚡ Supported Engine Variants

This repository provides automated build pipelines and release artifacts for three distinct ROCm inference engines:

flowchart TD
    subgraph Engines ["Inference Engine Variants"]
        E1["1. ROCmFPX Upstream (charlie12345/ROCmFPX)<br/>Official upstream engine • Active llama.cpp tracking • Multi-GPU HIP/MMQ"]
        E2["2. Ciru-AI ROCmFPX (ciru-ai/ROCmFPX)<br/>Ciru research fork • DualView Q7/Q8 • PromptForge • Kairic Edge profiles"]
        E3["3. q38rocm (julianmb/q38rocm)<br/>Dedicated Qwen 3.8 27B Strix Halo stack • 36 tok/s • MTP Speculative Decoding"]
    end

Loading

1. Upstream ROCmFPX (charlie12345/ROCmFPX)

The canonical, actively maintained upstream ROCmFPX project created by Charlie.

• Frequently synchronized with upstream llama.cpp.
• Multi-architecture HIP acceleration for AMD RDNA2, RDNA3, RDNA3.5, RDNA4, and CDNA.
• Standard high-performance MMQ/MMVQ dispatch.
• Homebrew Formula: brew install rocmfpx

2. Ciru-AI ROCmFPX (ciru-ai/ROCmFPX)

Ciru's specialized downstream research fork featuring low-bit quantization layouts:

• 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.
• ActiveFPX PromptForge: Prompt-specialized runtime featuring fused projections on Strix Halo (gfx1151).
• Certified Profiles: kairic-edge and promptforge.
• Homebrew Formula: brew install ciru-rocmfpx

3. q38rocm (julianmb/q38rocm)

Julian's dedicated deployment stack for Qwen 3.8 27B on AMD Strix Halo (Ryzen AI Max+ 395 / Radeon 8060S) APUs:

• Sustained 30.56 – 36.04 tok/s generation throughput via MTP (Multi-Token Prediction) Speculative Decoding (K=4..6).
• Asymmetric TurboQuant KV Cache (-ctk q8_0 -ctv turbo4): Compresses 262K context RAM from 61.4 GB to 20.08 GB.
• Mesa RADV Wave64 cooperative matrices (KHR_coopmat).
• Homebrew Formula: brew install q38rocm

Important

⚡ Ready to Run — ROCm™ 7 Built-in: All binaries include complete ROCm 7 runtime libraries, hipBLAS, rocBLAS, and hipBLASLt kernels with portable $ORIGIN RPATHs. 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, RX 7900 XTX / XT / GRE, RX 7800 XT, Radeon 780M / 760M
gfx103X	RDNA2 dGPUs & Handhelds	Steam Deck (Van Gogh), Radeon 680M, RX 6950 XT / 6800 XT / 6700 XT
gfx90a	CDNA2 Accelerators	AMD Instinct MI210, MI250, MI250X
gfx908	CDNA1 Accelerators	AMD Instinct MI100

---

🍺 Homebrew Tap Installation

Install directly via the Heretek-AI/homebrew-tap:

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

# 2. Install Upstream ROCmFPX (charlie12345/ROCmFPX)
brew install rocmfpx

# Or install Ciru-AI ROCmFPX (DualView Q7 / PromptForge / Kairic Edge)
brew install ciru-rocmfpx

# Or install dedicated q38rocm (Qwen 3.8 27B @ 36 tok/s on Strix Halo)
brew install q38rocm

---

🧠 Running Qwen 3.8 27B with 1M YaRN Context Extension

To run Qwen 3.8 27B with a 1,048,576 token (1M) context window on AMD Strix Halo / ROCm using YaRN RoPE interpolation and 4-bit KV cache:

1. Using the Convenience Script

./scripts/run_yarn_1m.sh -m /path/to/Qwen3.8-27B-ROCmFP4-STRIX_LEAN.gguf --mmproj /path/to/mmproj-F16.gguf

2. Direct llama-server CLI Invocation

HIP_VISIBLE_DEVICES=1 /home/linuxbrew/.linuxbrew/opt/q38rocm/bin/llama-server \
  --host 0.0.0.0 \
  --port 8800 \
  -m /home/ronin/Projects/models/Qwen-3.8-27B-ROCmFP4-FAST-GGUF/Qwen3.8-27B-ROCmFP4-STRIX_LEAN.gguf \
  --mmproj /home/ronin/Projects/models/Qwen-3.8-27B-ROCmFP4-FAST-GGUF/mmproj-F16.gguf \
  --image-min-tokens 1024 \
  -ngl 99 \
  -fit off \
  -np 1 \
  -c 1048576 \
  --override-kv qwen2.context_length=int:1048576,qwen2vl.context_length=int:1048576,qwen3.context_length=int:1048576 \
  --rope-scaling yarn \
  --rope-scale 4.0 \
  --yarn-orig-ctx 262144 \
  --yarn-ext-factor -1 \
  --yarn-attn-factor 1.0 \
  --yarn-beta-slow 1 \
  --yarn-beta-fast 32 \
  --cache-type-k q4_0 \
  --cache-type-v q4_0 \
  -fa on \
  -b 2048 \
  -ub 512

Parameter Notes:

• -c 1048576: Allocates the 1M token context buffer.
• --override-kv qwen2.context_length=int:1048576...: Overrides GGUF metadata n_ctx_train so llama-server initializes the full 1M slot context without capping at 262K.
• -fit off: Disables automatic VRAM fitting when full layers (-ngl 99) are explicitly requested.
• --rope-scaling yarn --rope-scale 4.0: Scales frequencies 4× from Qwen 3.8's native 262,144 base window (--yarn-orig-ctx 262144).
• --cache-type-k q4_0 --cache-type-v q4_0: 4-bit KV cache reduces memory footprint to ~40 GB, easily fitting within 128 GB unified memory on AMD Strix Halo (Ryzen AI Max+ 395).
• --image-min-tokens 1024: Ensures required vision grounding tokens for Qwen-VL multimodal inputs.

---

🚀 CI Workflows

• .github/workflows/build-rocmfpx.yml: Multi-OS, multi-GPU matrix builder for canonical upstream charlie12345/ROCmFPX.
• .github/workflows/build-rocmfpx-profile.yml: On-demand certified builder for ciru-ai/ROCmFPX (kairic-edge, promptforge, DualView).
• .github/workflows/build-q38rocm.yml: Dedicated Strix Halo gfx1151 builder for q38rocm.

---

📄 License & Attribution

• ROCmFPX (Upstream): Developed by Charlie (charlie12345/ROCmFPX) under the MIT License.
• ROCmFPX (Ciru Fork): Developed by Ciru (ciru-ai/ROCmFPX) under the MIT License.
• q38rocm: Developed by Julian (julianmb/q38rocm) under the Apache 2.0 License.
• llama.cpp: Developed by Georgi Gerganov and contributors under the MIT License.