v0.3.15

What's New

• PyTorch 2.10.0 support — Build matrix and fat wheel now include PyTorch 2.10.0
• CUDA 13.0 (cu130) support — Linux, with version-specific wheels for all platforms
• Fixed Windows builds — Removed torch/extension.h from CUDA sources to avoid MSVC header bugs
• Updated fat wheel to build cu126/cu128/cu130 extensions with PyTorch 2.10.0 on Linux

Supported Matrix

PyTorch	CUDA
2.5.1	cu118, cu121, cu124
2.6.0	cu118, cu124, cu126
2.7.1	cu118, cu126, cu128
2.8.0	cu126, cu128
2.9.1	cu126, cu128, cu130 (Linux)
2.10.0	cu126, cu128, cu130 (Linux)

Note: PyTorch 2.9.1/2.10.0 have Windows compilation bugs affecting CUDA extension builds. Windows fat wheel uses PyTorch 2.8.0 for cu126/cu128.

Full Changelog: v0.3.14...v0.3.15

v0.3.14

What's New

• PyTorch 2.9.1 support (Linux) - Updated fat wheel to latest stable
• Windows fix - Use PyTorch 2.8.0 for Windows builds (2.9.x has build bug)

Fat Wheel CUDA Variants

CUDA	Linux	Windows
cu118	2.7.1	2.7.1
cu124	2.6.0	2.6.0
cu126	2.9.1	2.8.0
cu128	2.9.1	2.8.0

Installation

pip install fussim

The fat wheel automatically detects your CUDA version at runtime.

v0.3.11

Important: Fat Wheel Strategy Change

PyTorch's C++ ABI is NOT forward compatible between minor versions. Extensions built with older PyTorch will not load with newer PyTorch.

Fat Wheel (pip install fussim)

Targets latest PyTorch users:

CUDA	PyTorch	Target Users
cu118	2.7.1	PyTorch 2.7.x
cu124	2.6.0	PyTorch 2.6.x
cu126	2.8.0	PyTorch 2.8.x
cu128	2.8.0	PyTorch 2.8.x

Older PyTorch Versions

Users on older PyTorch should use torch-specific wheels:

pip install fussim --extra-index-url https://opsiclear.github.io/fussim/whl

This provides wheels for specific PyTorch versions (2.5.1, 2.6.0, 2.7.1, etc.)

v0.3.10

Changes

Fixed Python 3.13 + cu124 exclusion

• cu118/cu124 use PyTorch 2.5.1 which doesn't support Python 3.13
• cu126/cu128 use PyTorch 2.6.0/2.7.0 which DO support Python 3.13

Fat Wheel Python 3.13 Support

CUDA	PyTorch	Python 3.13
cu118	2.5.1	❌ excluded
cu124	2.5.1	❌ excluded
cu126	2.6.0	✅ supported
cu128	2.7.0	✅ supported

Installation

pip install fussim

v0.3.9

Changes

Removed CUDA 13.0 (cu130)

• CUDA 13.0 toolkit not yet available in CI
• PyTorch cu130 wheels not yet widely available

Correct PyTorch Version Matrix

Fat wheel extensions built with oldest compatible PyTorch for ABI compatibility:

CUDA	PyTorch	Supports
cu118	2.5.1	PyTorch 2.5.x+
cu124	2.5.1	PyTorch 2.5.x+
cu126	2.6.0	PyTorch 2.6.x+
cu128	2.7.0	PyTorch 2.7.x+

Installation

pip install fussim

v0.3.8

Changes

Fixed CUDA/PyTorch Version Mappings

Build extensions with the oldest PyTorch version for each CUDA variant to ensure forward ABI compatibility:

CUDA Variant	PyTorch Version	Supports
cu118	2.5.1	2.5.x supports cu118, cu121, cu124
cu124	2.5.1	2.5.x supports cu118, cu121, cu124
cu126	2.6.0	2.6.x supports cu118, cu124, cu126
cu128	2.7.0	2.7.x supports cu118, cu126, cu128
cu130	2.9.0	2.9.x supports cu126, cu128, cu130

Added CUDA 13.0 Support

• New cu130 variant for PyTorch 2.9.0+ with CUDA 13.0

Installation

pip install fussim

Source: https://pytorch.org/get-started/previous-versions/

v0.3.7

Changes

• Fix ABI compatibility: Build CUDA extensions with the oldest PyTorch version for each CUDA variant
  • cu118: PyTorch 2.5.1
  • cu124: PyTorch 2.5.1
  • cu126: PyTorch 2.6.0 (was 2.8.0)
  • cu128: PyTorch 2.8.0

This ensures extensions work with all newer PyTorch versions (forward compatibility).

Installation

pip install fussim

v0.3.6

Changes

• Fix PyPI upload by skipping sdist (license-file metadata issue)
• Wheels-only distribution (CUDA extensions require pre-built binaries)

Installation

pip install fussim

For specific PyTorch/CUDA versions, use the custom index:

pip install fussim --extra-index-url https://opsiclear.github.io/fussim/whl

v0.3.5

Changes

This release fixes YAML syntax errors that caused v0.3.4 workflow to fail.

Bug Fixes

• Fixed YAML syntax errors from heredocs with content at column 0
• Replaced bash heredocs with printf for wheel metadata files
• Converted Python heredoc to inline python -c command
• Fixed METADATA indentation issue from v0.3.3
• Fixed Windows wheel ZIP corruption (mis-matched data size)

Improvements

• Added wheel validation before PyPI upload
• Enabled verbose logging in PyPI publish
• Improved wheel creation for cross-platform compatibility

Installation

pip install fussim

Pre-built wheels include all CUDA versions (11.8, 12.4, 12.6, 12.8) - the correct one is auto-detected at runtime.

v0.3.4

Changes

This release fixes the PyPI upload issue from v0.3.3 (which was partially uploaded and is unusable).

Bug Fixes

• Fixed METADATA files having invalid leading spaces from heredoc indentation
• Fixed Windows wheel ZIP corruption (mis-matched data size error)
• Added Description-Content-Type header to eliminate metadata warnings

Improvements

• Improved wheel ZIP creation with explicit ZipInfo settings for cross-platform reliability
• Added built-in wheel validation after creation
• Added twine check validation before PyPI upload
• Enabled verbose logging in PyPI publish for better debugging

Installation

pip install fussim

Pre-built wheels include all CUDA versions (11.8, 12.4, 12.6, 12.8) - the correct one is auto-detected at runtime.