| Index | index by Group | index by Distribution | index by Vendor | index by creation date | index by Name | Mirrors | Help |
The search service can find package by either name (apache), provides(webserver), absolute file names (/usr/bin/apache), binaries (gprof) or shared libraries (libXm.so.2) in standard path. It does not support multiple arguments yet...
The System and Arch are optional added filters, for example System could be "redhat", "redhat-7.2", "mandrake" or "gnome", Arch could be "i386" or "src", etc. depending on your system.
Tensors and dynamic neural networks in Python. Compiled from the official pytorch-v2.15.0 source tarball (not a PyPI wheel). 2.15 makes the rewritten NCCL backend the default (nccl-legacy, or TORCH_DIST_USE_NCCL2=0, rolls that back for this release), adds one-sided distributed transports, and supports torch.compile on Python 3.15. Official ROCm wheels are 7.14 and 10.0 and drop gfx906; this package is built against TheRock 10.1 (HIP 7.16) and still includes gfx906. GPU: * ROCm/HIP — AMD (gfx906, gfx908, gfx90a, gfx942, gfx1030, gfx1100–1102, gfx1200–1201). This is torch.cuda on Radeon / Instinct (ComfyUI, Krita AI). * Vulkan — in-tree ATen compute backend (USE_VULKAN). The desktop torch.device("vulkan") backend is the separate python-torchvulkan package. CUDA is off. Flash-attention/AOTriton is off (it downloads prebuilt images; ABF has no network). LTO is off (libtorch link OOMs with full LTO on 64G builders). Distributed is on (Gloo + RCCL + MPI). Kineto (ROCm profiler) is on. Bundled CK GEMM/SDPA/MSLK are on (Clang 23 raw_buffer_* signedness patched). hipFile is on.
| Package | Summary | Distribution | Download |
| python-torch-2.15.0-1.aarch64.html | PyTorch machine learning framework | OpenMandriva Cooker for aarch64 | python-torch-2.15.0-1.aarch64.rpm |
| python-torch-2.15.0-1.x86_64.html | PyTorch machine learning framework | OpenMandriva Cooker for x86_64 | python-torch-2.15.0-1.x86_64.rpm |
Generated by rpm2html 1.6