| --- |
| license: mit |
| tags: |
| - llama-cpp |
| - rocm |
| - amd |
| - gfx1201 |
| - windows |
| - inference-engine |
| - experimental |
| --- |
| |
| # ROCmFPX Runtime for Windows gfx1201 |
|
|
| Windows x64 inference runtime for ROCmFPX GGUF models, tested on AMD Radeon RX 9070 (`gfx1201`). Community build based on [llama.cpp](https://github.com/ggml-org/llama.cpp) and [ROCmFPX](https://github.com/charlie12345/ROCmFPX). |
|
|
| ## Downloads |
|
|
| | Version | Runtime and documentation | |
| | --- | --- | |
| | **20261022-01** | [Integrated llama.cpp release](https://huggingface.co/Roland2531/llama.cpp-ROCmFPX-gfx1201-Windows-x64/tree/main/releases/20261022-01) | |
| | 20261021-01 | [Previous runtime](https://huggingface.co/Roland2531/llama.cpp-ROCmFPX-gfx1201-Windows-x64/tree/main/releases/20261021-01) · [Compatible patch add-on](https://huggingface.co/Roland2531/llama.cpp-ROCmFPX-gfx1201-Windows-x64/tree/main/releases/20261021-01/fixes) | |
|
|
| Download the [matching Ornith Q6 IO + A8 model](https://huggingface.co/Roland2531/Ornith-1.5-9B-uncensored-ROCmI4-Q6IO-A8) separately. This repository contains the runtime; model weights are in the model repository. |
|
|
| ## Requirements |
|
|
| - Windows x64, a compatible AMD driver and Microsoft Visual C++ x64 runtime. |
| - GPU validation covers RX 9070 / gfx1201. Other GPUs and operating systems are unverified. |
| - Extract a complete runtime ZIP, including AMD DLLs, library data and sidecar files. Keep all files from the same version together. |
| - Follow the selected version's instructions; the releases have different archive layouts and DLL interfaces. Patch add-ons apply only to their stated version. |
|
|
| ## Capabilities and limits |
|
|
| The newer release supports the tested Q6 IO + W4A8 model, short-prefix/RAM caching, native tool grammar and projector lifecycle options. Its notes include absolute prefill/decode/latency and memory measurements. |
|
|
| Valid tool JSON does not guarantee correct instructions: exact wording and silence requests can still fail. The newer release is not a universal speed or model-quality upgrade. Real vision, Viewer/audio integration and full-context coexistence are not validated by its published tests. |
|
|
| ## Licenses |
|
|
| Core code is MIT. Bundled dependencies and model weights retain their own terms. See [LICENSE](https://huggingface.co/Roland2531/llama.cpp-ROCmFPX-gfx1201-Windows-x64/blob/main/LICENSE), [third-party notices](https://huggingface.co/Roland2531/llama.cpp-ROCmFPX-gfx1201-Windows-x64/blob/main/THIRD_PARTY_NOTICES.md) and package license directories. |
|
|
