Construction de CTranslate2 avec CUDA pour arm64 : roue, paquets Debian, CI Gitea
This commit is contained in:
commit
402f97df2a
17 files changed
+533
No files matched your search
Vendored
+40
@@ -0,0 +1,40 @@
|
||||
Source: ctranslate2
|
||||
Section: libs
|
||||
Priority: optional
|
||||
Maintainer: Eric Lemesre <eric.lemesre@gmail.com>
|
||||
Build-Depends: debhelper-compat (= 13),
|
||||
cmake (>= 3.15),
|
||||
ninja-build,
|
||||
libopenblas-dev
|
||||
Standards-Version: 4.6.2
|
||||
Homepage: https://github.com/OpenNMT/CTranslate2
|
||||
Rules-Requires-Root: no
|
||||
|
||||
Package: libctranslate2-4
|
||||
Architecture: arm64
|
||||
Multi-Arch: same
|
||||
Depends: ${shlibs:Depends}, ${misc:Depends}
|
||||
Recommends: libcublas-13-0, libcudnn9-cuda-13
|
||||
Description: fast inference engine for Transformer models (CUDA 13, arm64)
|
||||
CTranslate2 is a C++ and Python library for efficient inference with
|
||||
Transformer models (Whisper, NLLB, T5, Llama, ...), with weight quantization
|
||||
and CPU/GPU execution.
|
||||
.
|
||||
This build enables the CUDA backend for Linux arm64 (Jetson Orin, Grace
|
||||
Hopper, DGX Spark GB10), which the upstream project does not publish for
|
||||
this architecture. CUDA libraries are loaded at run time (cuBLAS 13,
|
||||
cuDNN 9) from the CUDA toolkit or from the nvidia-* pip packages; without
|
||||
them the library still works on CPU (OpenBLAS + Ruy).
|
||||
|
||||
Package: libctranslate2-dev
|
||||
Architecture: arm64
|
||||
Multi-Arch: same
|
||||
Section: libdevel
|
||||
Depends: libctranslate2-4 (= ${binary:Version}), ${misc:Depends}
|
||||
Description: fast inference engine for Transformer models - development files
|
||||
CTranslate2 is a C++ and Python library for efficient inference with
|
||||
Transformer models.
|
||||
.
|
||||
This package contains the headers and the unversioned shared library link
|
||||
needed to build programs against libctranslate2 (C++ or FFI bindings such
|
||||
as the Rust crate ct2rs).
|
||||
Reference in new issue
Block a user