whisper.cpp-rocm

Pre-built releases of whisper.cpp with full AMD hardware acceleration — ROCm™ GPU, Vulkan GPU, RyzenAI NPU, and optimised CPU builds — for Windows and Linux.

Releases track upstream whisper.cpp exactly: every time upstream publishes a new version, our automated pipeline syncs, builds all backends, and publishes a matching release within 24 hours. No manual steps. No lag.

Important

No ROCm installation required. All ROCm and Vulkan runtime libraries are bundled inside every release archive. Download, extract, and run.

Note

This project is maintained by the Lemonade SDK team. Our primary focus is seamless integration with Lemonade and similar AMD-optimised AI applications. We welcome collaborations and contributions that advance AMD whisper.cpp support.

---

🎯 Supported Devices

ROCm GPU

Architecture	Devices
gfx1151 — RDNA3.5 APU	Ryzen AI MAX+ Pro 395 (Strix Halo)
gfx1150 — RDNA3.5 APU	Ryzen AI 300 series (Strix Point)
gfx120X — RDNA4 dGPU	Radeon RX 9070 XT / 9070 / 9060 XT / 9060
gfx110X — RDNA3 dGPU & iGPU	RX 7900 XTX/XT/GRE, RX 7800 XT, RX 7700 XT, RX 7600 XT/7600; iGPU Radeon 780M / 760M / 740M

Vulkan GPU

Any GPU with a Vulkan 1.3-capable driver — AMD, NVIDIA, Intel. Covers iGPUs on all platforms where a Vulkan driver is present.

NPU — RyzenAI

Device	OS	Requirement
Ryzen AI 300 series (Strix Point / Strix Halo)	Windows only	NPU driver ≥ .280

CPU

Optimised CPU-only builds for x86-64. Windows and Linux. No GPU required.

---

📦 Downloads

All builds are self-contained — no separate driver or runtime installation needed (except the NPU driver for the NPU build).

ROCm — GPU Accelerated

GPU Target	Linux	Windows
gfx1151 (Ryzen AI MAX+ Pro 395)
gfx1150 (Ryzen AI 300)
gfx120X (RDNA4 dGPU)
gfx110X (RDNA3 dGPU & iGPU)

Vulkan — Cross-Vendor GPU

Linux	Windows

NPU — RyzenAI (Windows only)

Windows

Requires NPU driver ≥ .280 and a pre-compiled .rai encoder model from AMD's Hugging Face collection. Place the .rai file alongside your ggml-*.bin model — whisper-cli picks it up automatically.

macOS — Metal GPU

macOS (Apple Silicon)

CPU — No GPU Required

Linux	Windows

---

🧪 Quick Smoketest

1. Get a model

# Download the tiny.en model (~75 MB) for a fast smoke test
./models/download-ggml-model.sh tiny.en

# Or grab any ggml-*.bin from https://huggingface.co/ggerganov/whisper.cpp

2. Transcribe the bundled sample

# Linux
./whisper-cli -m models/ggml-tiny.en.bin -f samples/jfk.wav

# Windows
whisper-cli.exe -m models\ggml-tiny.en.bin -f samples\jfk.wav

Expected: a transcription of the JFK "Ask not what your country can do for you" excerpt.

3. Verify GPU is active (ROCm)

# At startup whisper-cli prints the backend in use — look for:
#   ggml_hip: using device ...
./whisper-cli -m models/ggml-tiny.en.bin -f samples/jfk.wav 2>&1 | grep -i "hip\|rocm\|device"

4. Verify NPU is active (VitisAI)

# Place the .rai encoder alongside the .bin model, then run normally.
# Look for this line in stdout:
#   whisper_vitisai_encode: Vitis AI model inference completed.
whisper-cli.exe -m models\ggml-tiny.en.bin -f samples\jfk.wav

5. Verify portability (Linux ROCm)

# ROCm runtime libs are bundled — RPATH should point to $ORIGIN (same dir as binary)
readelf -d whisper-cli | grep RPATH    # -> $ORIGIN
ldd whisper-cli | grep "not found"     # -> (empty — all deps resolved locally)

---

🔄 Release Cadence

Releases are fully automated and mirror upstream whisper.cpp releases with no manual steps:

upstream whisper.cpp releases vX.Y.Z
            |
            v  (detected within 24 h by daily sync job)
  sync.yml merges upstream into main, pushes tag vX.Y.Z
            |
            v  (tag push triggers build pipeline)
  build.yml builds all backend/OS combinations in parallel
            |
            v
  GitHub Release: "whisper.cpp vX.Y.Z — AMD Builds"
  with 13 artifacts across all backends and OS targets

Every release ships up to 14 artifacts:

whisper-{version}-linux-rocm-gfx1151.tar.gz
whisper-{version}-linux-rocm-gfx1150.tar.gz
whisper-{version}-linux-rocm-gfx120X.tar.gz
whisper-{version}-linux-rocm-gfx110X.tar.gz
whisper-{version}-windows-rocm-gfx1151.zip
whisper-{version}-windows-rocm-gfx1150.zip
whisper-{version}-windows-rocm-gfx120X.zip
whisper-{version}-windows-rocm-gfx110X.zip
whisper-{version}-linux-vulkan-x86_64.tar.gz
whisper-{version}-windows-vulkan-x64.zip
whisper-{version}-windows-npu-x64.zip         (may be absent if NPU runner offline)
whisper-{version}-linux-cpu-x86_64.tar.gz
whisper-{version}-windows-cpu-x64.zip
whisper-{version}-darwin-metal-arm64.tar.gz

Tip

Linux APU out of VRAM despite free memory (gfx1150 / gfx1151)? Add ttm.pages_limit=12582912 to your kernel command line (e.g. in GRUB), run update-grub, and reboot. See the TheRock FAQ for details.

---

🖥️ Local Builds (Windows)

Reproduce any CI build locally using the bundled PowerShell script. Produces identical artifacts to what CI publishes.

# Prerequisites: CMake, VS Build Tools 2022, 7-Zip, internet access

# CPU only (~2 min, no GPU needed)
.\scripts\local-build.ps1 -Backend cpu

# Vulkan — requires Vulkan SDK from https://vulkan.lunarg.com
.\scripts\local-build.ps1 -Backend vulkan

# ROCm for RDNA3 iGPU — downloads ROCm tarball (~2-4 GB, cached after first run)
.\scripts\local-build.ps1 -Backend rocm -GfxTarget gfx1151

# NPU — requires RyzenAI hardware + NPU driver >= .280
.\scripts\local-build.ps1 -Backend npu

# All backends, version-stamped artifacts placed in .\dist\
.\scripts\local-build.ps1 -Backend all -Version 1.8.4

---

📦 Dependencies

Bundled in every release (no installation needed)

Backend	What is included
ROCm	amdhip64, rocblas, hipblaslt + library data, LLVM runtime, all system deps; RPATH=$ORIGIN on Linux
Vulkan	SPIR-V shaders embedded at build time; links against system Vulkan loader
Metal	Uses macOS system Metal framework; no extra bundling needed
NPU	FlexML Runtime DLLs (flexmlrt/bin + flexmlrt/lib)
CPU	SDL2.dll included on Windows

Build-time only

Tool	Purpose
whisper.cpp	Upstream source
ROCm / TheRock	HIP compiler + GPU runtime (tarball, not installed globally)
FlexML Runtime	VitisAI NPU inference
Vulkan SDK	GLSL to SPIR-V shader compilation
CMake >= 3.21	Build system
Ninja	Fast build backend (ROCm builds)
VS Build Tools 2022	Windows MSVC toolchain

---

🏗️ Repository Structure

whisper.cpp-rocm/
├── .github/
│   └── workflows/
│       ├── build.yml           # All AMD backends — builds + publishes releases
│       └── sync.yml            # Daily upstream sync + auto-tagging
├── ci/
│   ├── resolve-rocm-version.sh    # Resolves AMD tarball URL for a given ROCm version
│   └── map-gpu-target.sh          # Maps gfx110X/gfx120X shorthands to specific arch lists
├── src/
│   └── vitisai/
│       ├── whisper-vitisai-encoder.h    # VitisAI NPU encoder C interface
│       └── whisper-vitisai-encoder.cpp  # FlexML runtime integration
├── scripts/
│   └── local-build.ps1         # Local Windows build script (mirrors CI jobs exactly)
├── ggml/                       # GGML library (all GPU backends live here)
├── src/                        # whisper.cpp source (VitisAI hooks added)
└── CMakeLists.txt              # Adds -DWHISPER_VITISAI option

---

📄 License

This project is licensed under the MIT License — see LICENSE for details.

whisper.cpp is copyright Georgi Gerganov and contributors — ggerganov/whisper.cpp. ROCm is copyright Advanced Micro Devices, Inc. VitisAI encoder copyright 2025 Advanced Micro Devices, Inc.