🚨 Gated delta net prefill was still walking tokens one by one.

RDNA3 now does 16-token chunks on WMMA. State stays in the F32 accumulators. Scalar gate, head dim 128, one sequence. KDA and other head dims fall back. Snapshot tail still uses the old kernel.

⚡ 772 µs → 610 µs per 512-token call ⚡ pp4096 on main, ub 1024: 850 → 882–884 (+3.7%). ub 512: 645–662 → 664–665 ⚡ with the other open HIP PRs, pp8192 ub 2048 +3.5%, ub 8192 +2–4% ✅ KL 0.016301 / 95.99% → 0.016307 / 95.96%

https://github.com/ROCmFPX/ROCmFPX/pull/45…

@Italianclownz @AIatAMD

And the Strix Halo testing continues…

ROCmFPX/ROCmFPX

Source: https://github.com/ROCmFPX/ROCmFPX

ROCmFPX llama.cpp

This public downstream tracks current upstream llama.cpp while developing the ROCmFPX weight-format family for AMD GPUs. Normal Vulkan support remains enabled; HIP and CPU are also supported build targets. Existing NVFP4 GGUF tensors can be loaded natively and remain bit-exact when an NVFP4 model is completed with the NVFP4 quantization preset.

Carlo Pasquale (Charlie12345) is the creator and founder of the ROCmFPX format family and of ROCmFP3, ROCmFP4, ROCmFP6, and ROCmFP8. See ROCmFPX documentation, NOTICE, and the same upstream MIT license terms with the ROCmFPX copyright line.

Contributors and history

ROCmFPX preserves the upstream llama.cpp lineage and the public legacy ROCmFPX lineage. Original commit authors and commit IDs remain reachable. The legacy lineage is connected by an ancestry-only merge whose source tree is identical to the current ROCmFPX tree, so legacy code does not replace the current implementation.

• Current ROCmFPX contributors
• Legacy ROCmFPX contributors
• Upstream llama.cpp contributors

GitHub contributor displays can lag behind repository history. The complete commit-level author record is also available with git shortlog -sne main.

The currently qualified formats are experimental. The on-disk layouts of ROCmFP2/3/4/6/8 are frozen for compatibility; new kernel and quantizer work must preserve their encoded sizes and semantics.

ROCmFP2 tensor type identity

Canonical dual-scale S40 ROCmFP2 uses GGUF tensor type 111. Its payload is still exactly 10 bytes per 32 weights (2.5 bpw); only the tensor type tag moved. Legacy type 107 is reserved as ambiguous because both dual-scale S40 and an affine code * scale - offset layout were emitted with that same ID and block size. ROCmFPX refuses type 107 instead of guessing and silently corrupting a model.

New quantizations write type 111 automatically. To audit an older file, run:

scripts/rocmfpx/retag-legacy-rocmfp2.py MODEL.gguf

Only when the file’s provenance confirms that it uses dual-scale S40, make a backup and retag its tensor headers in place:

scripts/rocmfpx/retag-legacy-rocmfp2.py MODEL.gguf \
  --layout s40-dual-scale-v1 --apply

The tool does not inspect or infer the layout because the two interpretations cannot be distinguished from the bytes. Do not use it on affine ROCmFP2 files.

Optional Charlie Vulkan plugin

The optional ROCmFPXVulkan extension preserves the proven Charlie-era ROCmFP4 Vulkan path as a separate backend. It does not replace or patch llama.cpp’s normal Vulkan0 backend. The extension is disabled by default and appears as ROCmFPXVulkan0 only when it is built and explicitly loaded.

Build both the normal Vulkan backend and the optional extension on Linux:

cmake -S . -B build-rocmfpx-vulkan -G Ninja \
  -DCMAKE_BUILD_TYPE=Release \
  -DBUILD_SHARED_LIBS=ON \
  -DGGML_VULKAN=ON \
  -DROCMFPX_VULKAN_PLUGIN=ON
cmake --build build-rocmfpx-vulkan \
  --target llama-cli rocmfpx-vulkan-plugin -j

Load the plugin and select its device:

export ROCMFPX_PLUGIN_PATH="$PWD/build-rocmfpx-vulkan/bin/rocmfpx-vulkan-plugin.so"
build-rocmfpx-vulkan/bin/llama-cli --list-devices
build-rocmfpx-vulkan/bin/llama-cli \
  -m /absolute/path/to/model.gguf \
  -dev ROCmFPXVulkan0 -ngl 999

Keep rocmfpx-vulkan-plugin.so and its sibling libggml-rocmfpx-vulkan.so together. Both files must come from the same build and ROCmFPX commit. --list-devices should show both Vulkan0 and ROCmFPXVulkan0; remove ROCMFPX_PLUGIN_PATH to return to the normal backend. See the extension guide for matched ROCmFP4 benchmarks and qualification limits.

Plugin system

ROCmFPX plugin ABI v1 lets trusted native shared libraries register a standard ggml backend or receive model-open and model-close notifications for PLE, KV checkpoint, repack-cache, and expert-streaming sidecars. Set ROCMFPX_PLUGIN_PATH to a plugin file or directory before starting a ROCmFPX tool. Use : between entries on Linux and macOS and ; on Windows. Directories are scanned once, non-recursively, in sorted order; the working directory is never searched automatically.

A minimal C plugin looks like this:

#define ROCMFPX_PLUGIN_BUILD
#include "rocmfpx-plugin.h"

static int on_load(void) {
    return 0;
}

static const struct rocmfpx_plugin_v1 plugin = {
    ROCMFPX_PLUGIN_ABI_VERSION,
    sizeof(struct rocmfpx_plugin_v1),
    "example-sidecar",
    "0.1.0",
    ROCMFPX_PLUGIN_CAP_SIDECAR,
    on_load,
    0,
    0,
    0,
};

ROCMFPX_PLUGIN_EXPORT const struct rocmfpx_plugin_v1 * rocmfpx_plugin_query(
        uint32_t host_abi_version,
        const struct rocmfpx_plugin_host_v1 * host) {
    if (host_abi_version != ROCMFPX_PLUGIN_ABI_VERSION || !host ||
        host->abi_version != ROCMFPX_PLUGIN_ABI_VERSION) {
        return 0;
    }
    return &plugin;
}

Build and load it on Linux:

cc -shared -fPIC -I/path/to/ROCmFPX/include \
  example-sidecar.c -o example-sidecar.so
ROCMFPX_PLUGIN_PATH="$PWD/example-sidecar.so" \
  /path/to/ROCmFPX/build/bin/llama-cli --list-devices

Every plugin must export rocmfpx_plugin_query, validate the ABI, return a static size-versioned descriptor, and keep that descriptor and its strings alive until unload. A backend plugin copies backend_load and plugin_path during the query, then registers its matching ggml backend from on_load. A sidecar uses on_model_open and on_model_close and keys its state by model_id. ABI v1 provides discovery, backend registration, and lifecycle notifications; capability flags alone do not intercept tensors or token generation. Plugins run as native code with the same permissions as ROCmFPX, so load only libraries you trust. See the complete plugin and sidecar ABI guide and the tested rocmfpx-test-plugin example.

Windows AMD multi-GPU bridge

Windows users with two AMD GPUs can evaluate Charlie12345’s external Windows AMD Multi-GPU Bridge. For ROCmFPX and llama.cpp, use its dedicated Windows installation guide, not the separate vLLM adapter instructions.

The bridge keeps the ROCmFPX source tree unchanged. It supplies an external roc::rccl CMake package to the existing HIP collective interface, so build ROCmFPX with GGML_HIP_RCCL=ON, point rccl_DIR at the installed bridge, and use the bridge’s run-with-llama-plugin.ps1 launcher. The tested Windows path also requires GGML_CUDA_NO_PEER_COPY=ON; follow the external guide for the matching ROCm version, GPU target, DLL layout, health probes, and model-specific launch flags.

Despite the launcher’s name, this transport is not loaded through ROCMFPX_PLUGIN_PATH and is not a ROCmFPX ABI v1 plugin. ABI v1 can register a ggml backend or receive model lifecycle notifications, but it cannot inject a link-time RCCL provider or intercept collectives. Pointing ROCMFPX_PLUGIN_PATH at rccl.dll will therefore do nothing. The current external-package design is the upstream-safe integration: update ROCmFPX normally, then configure a fresh HIP build against the bridge package.

The bridge is experimental, source-only, and currently qualified only on the hardware and revisions listed by its maintainers. Validate its small parity and transport probes before loading a large model; a selectable GPU target is not the same as a runtime-qualified configuration.

llama.cpp

Quick start

A few options to get llama.cpp installed on your machine:

• Visit https://llama.app and follow the instructions
• Run with Docker - see our Docker documentation
• Download pre-built binaries from the releases page
• Build from source by cloning this repository - check out our build guide

Once installed:

# Download and run a model directly from Hugging Face
llama cli -hf ggml-org/Qwen3.5-0.8B-GGUF

# Launch OpenAI-compatible API server
llama serve -hf ggml-org/Qwen3.5-0.8B-GGUF

VLM session with llama cli

Built-in web UI against llama serve

Description

The main goal of llama.cpp is to enable LLM (and VLM) inference with minimal setup and state-of-the-art performance on a wide range of hardware - locally and in the cloud.

• Plain C/C++ implementation without any dependencies
• Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
• AVX, AVX2, AVX512 and AMX support for x86 architectures
• RVV, ZVFH, ZFH, ZICBOP and ZIHINTPAUSE support for RISC-V architectures
• 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
• Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP and Moore Threads GPUs via MUSA)
• Vulkan and SYCL backend support
• CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity

The llama.cpp project is build on top of the ggml library.

Supported backends

Documentation

Tools

• cli
• completion
• server
• GBNF grammars

Development

• How to build
• Running on Docker
• Build on Android
• Multi-GPU usage
• Performance troubleshooting
• GGML tips & tricks
• XCFramework
• Completions
• Models
• Release process

Contributing

• Contributors can open PRs
• Collaborators will be invited based on contributions
• Maintainers can push to branches in the llama.cpp repo and merge PRs into the master branch
• Any help with managing issues, PRs and projects is very appreciated!
• Read the CONTRIBUTING.md for more information

Acknowledgements

• yhirose/cpp-httplib - Single-header HTTP server, used by llama-server - MIT license
• nothings/stb - Single-header image format decoder, used by multimodal subsystem - Public domain
• nlohmann/json - Single-header JSON library, used by various tools/examples - MIT License
• mackron/miniaudio - Single-header audio format decoder, used by multimodal subsystem - Public domain
• sheredom/subprocess.h - Single-header process launching solution for C and C++ - Public domain