The HIP patch already targets the ROCm 7 hipBLAS API (HIPBLAS_COMPUTE_*,
hipDataType, HIP_R_*). The sed block was explicitly downgrading those to
the deprecated ROCm 6.2 equivalents; removing it lets the patch build
natively against ROCm 7.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The builder had dev packages but the runtime stage was missing the .so
files, causing ImportError on libhipblas.so.2 at startup.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
ROCm 6.2 hipblasGemmEx uses hipblasDatatype_t for the compute type arg,
not hipblasComputeType_t (that came in ROCm 7). Map compute type and
HIPBLAS_COMPUTE_* constants down to the old hipblasDatatype_t/HIPBLAS_R_*
equivalents. Also fix HIPBLAS_R_16B (not 16BF) and add rocrand-dev.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add hiprand-dev and hipcub-dev for missing headers
- Patch cuda2hip_macros.hpp post-apply: downgrade hipDataType→hipblasDatatype_t
and HIP_R_* → HIPBLAS_R_* since ROCm 6.2 hipblasGemmEx still uses the
older hipblasDatatype_t signature (hipDataType wasn't wired into
hipblasGemmEx until ROCm 7.0)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Install hipblas-dev and miopen-hip-dev, and set CMAKE_PREFIX_PATH=/opt/rocm
so cmake can locate the hipblasConfig.cmake and MIOpen config files.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Multi-stage Dockerfile builds CTranslate2 v4.0.0 with HIP/ROCm support
targeting gfx1030/gfx1031 (RX 6000 series), then installs faster-whisper
and wyoming-faster-whisper on top for a Wyoming ASR server on port 10300.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
