docs: add Windows ROCm/HIP build guide#2041
Open
T0nd3 wants to merge 2 commits into
Open
Conversation
Adds a step-by-step guide for building CTranslate2 with AMD GPU support (ROCm/HIP) on Windows, validated on RX 7900 XTX (gfx1100) with ROCm 7.2. Changes: - docs/building_rocm_windows.md: new detailed guide covering all prerequisites (VS Build Tools, ROCm wheels, Intel MKL, oneDNN), CMake configuration, C++ and Python builds, and known limitations - docs/hardware_support.md: add AMD/ROCm GPU section - docs/installation.md: link to ROCm guide from pip and build sections - docs/index.rst: add building_rocm_windows to the Get started toctree
6 tasks
…ible Issue OpenNMT#2016 reports that bundling CTranslate2 inside a PyInstaller `.exe` breaks when the end user installs the standalone AMD HIP SDK Installer: the wheel is linked against `hipblas.dll` (the name in ROCm 7.2, which is what the rocm-sdk pip wheels ship), but the latest publicly available AMD HIP SDK Installer is still on 7.1.1, where the same library is named `libhipblas.dll`. The Windows dynamic loader can't find the symbol the wheel asks for and `import ctranslate2` fails with a generic "module not found" error. Add a `Why not the official AMD HIP SDK Installer?` subsection right under the existing "ROCm via Python wheels" instructions that: - documents the version skew and the DLL name change, - explicitly recommends the Python wheels for CTranslate2, - explains the two options for users who need the .exe installer (build against 7.1, or wait for AMD to ship a 7.2 installer). The original wording already said "this method does not require the AMD HIP SDK installer", but didn't make it obvious that mixing the two is actively broken — which is exactly what the issue reporter ran into.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Motivation
ROCm support was merged in v4.7.0 (February 2026) but no documentation existed for building from source on Windows. The Windows build has several non-obvious requirements that differ from the Linux CI script:
rc.exewhen Clang is the compilertar.exe -Cis unreliable on Windows; Pythontarfileis the safe alternativeos.add_dll_directorymust be called before importing the module at runtimeTest plan
get_cuda_device_count() == 1, all compute types (float16, bfloat16, int8) available