# whisper.cpp-rocm
**Repository Path**: taiyi-one/whisper.cpp-rocm
## Basic Information
- **Project Name**: whisper.cpp-rocm
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: MIT
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No
## Statistics
- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-08-09
- **Last Updated**: 2026-08-09
## Categories & Tags
**Categories**: Uncategorized
**Tags**: None
## README
# whisper.cpp-rocm
Pre-built releases of **[whisper.cpp](https://github.com/ggerganov/whisper.cpp)** with full AMD hardware acceleration β **ROCmβ’ GPU**, **Vulkan GPU**, **RyzenAI NPU**, and optimised **CPU** builds β 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](https://github.com/lemonade-sdk/lemonade) 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.
---
## π― Supported Devices
### ROCm GPU
| Architecture | Devices |
|---|---|
| **gfx1151** β RDNA3.5 APU | Ryzen AI MAX+ Pro 395 (Strix Halo) |
| **gfx1150** β RDNA3.5 APU | Ryzen AI 300 series (Strix Point) |
| **gfx120X** β RDNA4 dGPU | Radeon RX 9070 XT / 9070 / 9060 XT / 9060 |
| **gfx110X** β 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 β AMD, NVIDIA, Intel. Covers iGPUs on all platforms where a Vulkan driver is present.
### NPU β RyzenAI
| Device | OS | Requirement |
|---|---|---|
| Ryzen AI 300 series (Strix Point / Strix Halo) | Windows only | NPU driver β₯ `.280` |
### CPU
Optimised CPU-only builds for x86-64. Windows and Linux. No GPU required.
---
## π¦ Downloads
All builds are self-contained β no separate driver or runtime installation needed (except the NPU driver for the NPU build).
### ROCm β GPU Accelerated
| GPU Target | Linux | Windows |
|---|---|---|
| **gfx1151** (Ryzen AI MAX+ Pro 395) | [](https://github.com/lemonade-sdk/whisper.cpp-rocm/releases/download/v1.8.4/whisper-v1.8.4-release-linux-rocm-gfx1151.tar.gz) | [](https://github.com/lemonade-sdk/whisper.cpp-rocm/releases/download/v1.8.4/whisper-v1.8.4-release-windows-rocm-gfx1151.zip) |
| **gfx1150** (Ryzen AI 300) | [](https://github.com/lemonade-sdk/whisper.cpp-rocm/releases/download/v1.8.4/whisper-v1.8.4-release-linux-rocm-gfx1150.tar.gz) | [](https://github.com/lemonade-sdk/whisper.cpp-rocm/releases/download/v1.8.4/whisper-v1.8.4-release-windows-rocm-gfx1150.zip) |
| **gfx120X** (RDNA4 dGPU) | [](https://github.com/lemonade-sdk/whisper.cpp-rocm/releases/download/v1.8.4/whisper-v1.8.4-release-linux-rocm-gfx120X.tar.gz) | [](https://github.com/lemonade-sdk/whisper.cpp-rocm/releases/download/v1.8.4/whisper-v1.8.4-release-windows-rocm-gfx120X.zip) |
| **gfx110X** (RDNA3 dGPU & iGPU) | [](https://github.com/lemonade-sdk/whisper.cpp-rocm/releases/download/v1.8.4/whisper-v1.8.4-release-linux-rocm-gfx110X.tar.gz) | [](https://github.com/lemonade-sdk/whisper.cpp-rocm/releases/download/v1.8.4/whisper-v1.8.4-release-windows-rocm-gfx110X.zip) |
### Vulkan β Cross-Vendor GPU
| Linux | Windows |
|---|---|
| [](https://github.com/lemonade-sdk/whisper.cpp-rocm/releases/download/v1.8.4/whisper-v1.8.4-release-linux-vulkan-x86_64.tar.gz) | [](https://github.com/lemonade-sdk/whisper.cpp-rocm/releases/download/v1.8.4/whisper-v1.8.4-release-windows-vulkan-x64.zip) |
### NPU β RyzenAI (Windows only)
| Windows |
|---|
| [-red?logo=amd&logoColor=white)](https://github.com/lemonade-sdk/whisper.cpp-rocm/releases/download/v1.8.4/whisper-v1.8.4-release-windows-npu-x64.zip) |
> Requires NPU driver β₯ `.280` and a pre-compiled `.rai` encoder model from [AMD's Hugging Face collection](https://huggingface.co/collections/amd/ryzen-ai-16-whisper-npu-optimized-onnx-models). Place the `.rai` file alongside your `ggml-*.bin` model β whisper-cli picks it up automatically.
### macOS β Metal GPU
| macOS (Apple Silicon) |
|---|
| [-lightgrey?logo=apple&logoColor=white)](https://github.com/lemonade-sdk/whisper.cpp-rocm/releases/download/v1.8.4/whisper-v1.8.4-release-darwin-metal-arm64.tar.gz) |
### CPU β No GPU Required
| Linux | Windows |
|---|---|
| [](https://github.com/lemonade-sdk/whisper.cpp-rocm/releases/download/v1.8.4/whisper-v1.8.4-release-linux-cpu-x86_64.tar.gz) | [](https://github.com/lemonade-sdk/whisper.cpp-rocm/releases/download/v1.8.4/whisper-v1.8.4-release-windows-cpu-x64.zip) |
---
## π§ͺ Quick Smoketest
### 1. Get a model
```bash
# 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
```bash
# 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)
```bash
# At startup whisper-cli prints the backend in use β 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)
```bash
# ROCm runtime libs are bundled β RPATH should point to $ORIGIN (same dir as binary)
readelf -d whisper-cli | grep RPATH # -> $ORIGIN
ldd whisper-cli | grep "not found" # -> (empty β all deps resolved locally)
```
---
## π 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 β 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](https://github.com/ROCm/TheRock/blob/main/docs/faq.md#gfx1151-strix-halo-specific-questions) for details.
---
## π₯οΈ Local Builds (Windows)
Reproduce any CI build locally using the bundled PowerShell script. Produces identical artifacts to what CI publishes.
```powershell
# Prerequisites: CMake, VS Build Tools 2022, 7-Zip, internet access
# CPU only (~2 min, no GPU needed)
.\scripts\local-build.ps1 -Backend cpu
# Vulkan β requires Vulkan SDK from https://vulkan.lunarg.com
.\scripts\local-build.ps1 -Backend vulkan
# ROCm for RDNA3 iGPU β downloads ROCm tarball (~2-4 GB, cached after first run)
.\scripts\local-build.ps1 -Backend rocm -GfxTarget gfx1151
# NPU β 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
```
---
## π¦ 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](https://github.com/ggerganov/whisper.cpp) | Upstream source |
| [ROCm / TheRock](https://github.com/ROCm/TheRock) | HIP compiler + GPU runtime (tarball, not installed globally) |
| [FlexML Runtime](https://github.com/lemonade-sdk/whisper.cpp/releases/tag/deps) | VitisAI NPU inference |
| [Vulkan SDK](https://vulkan.lunarg.com/sdk/home) | GLSL to SPIR-V shader compilation |
| [CMake >= 3.21](https://cmake.org/) | Build system |
| [Ninja](https://ninja-build.org/) | Fast build backend (ROCm builds) |
| [VS Build Tools 2022](https://visualstudio.microsoft.com/downloads/#build-tools-for-visual-studio-2022) | Windows MSVC toolchain |
---
## ποΈ Repository Structure
```
whisper.cpp-rocm/
βββ .github/
β βββ workflows/
β βββ build.yml # All AMD backends β builds + publishes releases
β βββ sync.yml # Daily upstream sync + auto-tagging
βββ ci/
β βββ resolve-rocm-version.sh # Resolves AMD tarball URL for a given ROCm version
β βββ map-gpu-target.sh # Maps gfx110X/gfx120X shorthands to specific arch lists
βββ src/
β βββ vitisai/
β βββ whisper-vitisai-encoder.h # VitisAI NPU encoder C interface
β βββ whisper-vitisai-encoder.cpp # FlexML runtime integration
βββ scripts/
β βββ local-build.ps1 # Local Windows build script (mirrors CI jobs exactly)
βββ ggml/ # GGML library (all GPU backends live here)
βββ src/ # whisper.cpp source (VitisAI hooks added)
βββ CMakeLists.txt # Adds -DWHISPER_VITISAI option
```
---
## π License
This project is licensed under the MIT License β see [LICENSE](LICENSE) for details.
whisper.cpp is copyright Georgi Gerganov and contributors β [ggerganov/whisper.cpp](https://github.com/ggerganov/whisper.cpp).
ROCm is copyright Advanced Micro Devices, Inc.
VitisAI encoder copyright 2025 Advanced Micro Devices, Inc.