llama.cpp (ROCmFPX fork) — Windows x64 build for Strix Halo
==========================================================

WHAT THIS IS
  A prebuilt Windows x64 binary distribution of the ROCmFPX fork of llama.cpp.
  This is the ONLY runtime that can load GGUF files using the ROCmFP4 / ROCmFPX
  data types (for example Q4_0_ROCMFP4_STRIX). Stock llama.cpp, LM Studio and
  Ollama cannot load those files at all — they will error out on unknown tensor
  types.

  It is published alongside the quantized model it was built for:
  https://huggingface.co/parsimo2010/Qwen3.8-27B-Q4_0_ROCMFP4_STRIX-GGUF


PROVENANCE
  fork     : https://github.com/charlie12345/ROCmFPX
  commit   : c49ebdbd5c9f01ec242369f9e7f7967855f80cba   (branch main)
  built    : 2026-09-18, Windows 11 x64
  compiler : MSVC 14.44.35207 (VS 2022 Build Tools) + AMD clang 21.0.0 (ROCm 7.1)
  target   : gfx1151  (AMD Strix Halo: Ryzen AI MAX 300-series + Radeon 8000S)

  IMPORTANT: gfx1151 device code is compiled into ggml-hip.dll. On any other AMD
  GPU the ROCm backend will not run — either use the Vulkan backend, or build the
  fork from source for your own GPU architecture.

BUILD FLAGS
  -DCMAKE_BUILD_TYPE=Release
  -DGGML_HIP=ON
  -DGGML_HIP_FORCE_MMQ=ON        <- required by the ROCmFP4 kernels, do not drop
  -DGGML_HIP_ROCWMMA_FATTN=OFF
  -DCMAKE_HIP_ARCHITECTURES=gfx1151
  -DGGML_VULKAN=ON
  -DLLAMA_BUILD_SERVER=ON


RUNTIME REQUIREMENTS
  These are NOT bundled — install them yourself.

  1. Windows 10 / 11, x64.

  2. Vulkan backend — easiest, needs almost nothing:
       Requires only vulkan-1.dll, which ships with a current AMD Adrenalin
       graphics driver. If your driver is reasonably new, this works with no
       extra installation. Use it by passing  -dev Vulkan0

  3. ROCm backend — faster prefill, needed for best performance:
       Requires the AMD ROCm HIP SDK, version 7.x.
       ggml-hip.dll imports libhipblas.dll and amdhip64_7.dll. Note that
       libhipblas.dll ships ONLY with the ROCm SDK — it is not part of the
       graphics driver, so having ROCm-related DLLs in C:\Windows\System32 is
       not sufficient. Install ROCm 7.x and put <ROCm install>\bin on PATH.
       Use it by passing  -dev ROCm0    (the "7" in amdhip64_7.dll means the
       installed ROCm must be 7.x, not 6.x.)

  4. Microsoft Visual C++ 2015-2022 Redistributable (x64), for MSVCP140.dll and
     VCRUNTIME140.dll. This is already present on most Windows systems.

  You do not need both backends — Vulkan alone is enough to run the model, and
  ROCm alone is enough as well. Builds with both are included so you can compare.


CONTENTS
  llama-server.exe       HTTP server, OpenAI-compatible API + embedded chat WebUI
  llama-cli.exe          interactive and one-shot command line
  llama-quantize.exe     quantization, including the ROCmFP4 types
  llama-imatrix.exe      importance matrix computation
  llama-bench.exe        benchmarking tool
  llama-perplexity.exe   perplexity evaluation
  test-backend-ops.exe   backend self-test — useful to confirm the build works
  ggml-hip.dll           ROCm/HIP backend        (~107 MB)
  ggml-vulkan.dll        Vulkan backend          (~88 MB)
  ggml-cpu.dll           CPU backend
  ggml-base.dll  ggml.dll  llama.dll  llama-common.dll  mtmd.dll
  LICENSE                MIT license (llama.cpp / the ggml authors)
  README.txt             this file
  SHA256SUMS             checksums for every file in this archive

  KEEP ALL THE DLLs IN THE SAME FOLDER AS THE .EXE FILES. The executables will
  not start if the matching DLLs are missing or in a different directory.


HOW TO RUN
  Open PowerShell in the folder you extracted this archive to.

  Chat UI / API server (tuned for a Strix Halo machine with 128 GB):

    .\llama-server.exe `
      -m <path to>\Qwen3.8-27B-Q4_0_ROCMFP4_STRIX.gguf `
      -dev ROCm0 -ngl 999 -fa on --jinja `
      -c 262144 -b 512 -ub 512 -t 16 `
      -ctk q8_0 -ctv q8_0 --cache-ram 10240 `
      --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-p-min 0.0 `
      --host 127.0.0.1 --port 8090

  Then open http://127.0.0.1:8090 for the built-in chat interface.

  If you do not have the ROCm SDK installed, swap -dev ROCm0 for -dev Vulkan0.
  Vulkan also needs no extra install, but prefill will be roughly 25-40% slower.

  Quick one-shot sanity check from the CLI:

    .\llama-cli.exe -m <path to>\Qwen3.8-27B-Q4_0_ROCMFP4_STRIX.gguf `
      -dev ROCm0 -ngl 999 -fa on -st `
      -p "Write a Python LRU cache with O(1) get and put." -n 128

  Verify the backends are working at all:

    .\test-backend-ops.exe -b ROCm0 -o MUL_MAT

  Full tuning notes, measured benchmarks and the reasoning behind every flag in
  the server command above are in the model card:
  https://huggingface.co/parsimo2010/Qwen3.8-27B-Q4_0_ROCMFP4_STRIX-GGUF


A NOTE ON WINDOWS SECURITY PROMPTS
  Because this archive contains .exe files downloaded from the internet, Windows
  may mark the extracted files as blocked and SmartScreen may warn you. If the
  executables refuse to run or Windows shows a security dialog:

    Right-click the .zip file -> Properties -> tick "Unblock" -> OK,
    and then extract it. (Unblocking before extracting is what matters.)


VERIFYING THE DOWNLOAD
  SHA256SUMS contains a SHA-256 hash for every file here. To check on Windows:

    Get-FileHash .\llama-server.exe -Algorithm SHA256

  or on Linux/macOS:

    sha256sum -c SHA256SUMS


LICENSE
  llama.cpp, ggml and this build are MIT licensed — see the LICENSE file, which
  must be kept with any redistribution.

  The ROCm / HIP runtime and the Vulkan driver are NOT included here. They are
  AMD's components and are covered by their own licenses; you install them
  yourself as described under RUNTIME REQUIREMENTS.