Skip to content
This repository was archived by the owner on Feb 18, 2026. It is now read-only.
This repository was archived by the owner on Feb 18, 2026. It is now read-only.

Patiently waiting for triton 3.2.0 for Windows now that torch 2.6.0 is released #50

Description

@BBC-Esq

Here's what torch is being built/compatible with pursuant to:

pytorch/.github/scripts/generate_binary_build_matrix.py (within the torch repository)

Here is a brief selection of the pip-installable libraries that torch 2.6.0 has been tested/compatible with:

+-------------------+------------+--------------+----------+----------+----------+---------+
| Torch Version     | cuda-nvrtc | cuda-runtime | cublas   | cufft     | cudnn    | triton |
+-------------------+------------+--------------+----------+-----------+----------+--------+
| 2.6.0 (CUDA 12.6) | 12.6.77    | 12.6.77      | 12.6.4.1 | 11.3.0.4  | 9.5.1.17 | 3.2.0  |
| 2.6.0 (CUDA 12.4) | 12.4.127   | 12.4.127     | 12.4.5.8 | 11.2.1.3  | 9.1.0.70 | 3.2.0  |
| 2.5.1 (CUDA 12.4) | 12.4.127   | 12.4.127     | 12.4.5.8 | 11.2.1.3  | 9.1.0.70 | 3.1.0  |
| 2.5.1 (CUDA 12.1) | 12.1.105   | 12.1.105     | 12.1.3.1 | 11.0.2.54 | 9.1.0.70 | 3.1.0  |
+-------------------+------------+--------------+----------+-----------+----------+--------+

Thus, apparently, PyTorch is building/testing with individual CUDA libraries originating from version 12.6.2 in their "cu126" wheels with on exception; their cublas version comes from CUDA 12.6.3.

And apparently torch 2.6.0 no longer builds wheels for CUDA 12.1

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions