pixtreme-infer-span v0.11.0

SPAN architecture module for pixtreme.infer, distributed as a Linux x86_64 platform wheel. The scale is read from the checkpoint.

Contents

• pixtreme_infer_span-0.11.0-py3-none-manylinux_2_28_x86_64.whl: fp16 cubins for sm86, sm89 and sm120 and one prebuilt TensorRT plugin library. Consumers need no C++ toolchain or SDK; engines are generated in the consumer environment.
• SHA256SUMS: checksum of the wheel. Verify with sha256sum -c SHA256SUMS.

Requirements

• Linux x86_64, glibc 2.28 or newer (for example RHEL 8 and 9, Ubuntu 22.04 and 24.04, Debian 12), Python 3.12 or newer
• TensorRT 11 (pip distribution tensorrt-cu13)
• Dependencies (lower bounds only): pixtreme-infer>=0.14.1, tensoracle>=1.1.0, safetensors>=0.4, numpy>=2.0
• Checkpoints are read from safetensors files. Converting a PyTorch .pth checkpoint is a separate, one-time step with python -m tensoracle.convert.

Changes since v0.10.0

• License: the package is released under the MIT License. The wheel's License-Expression is MIT AND Apache-2.0, and dist-info/licenses/ carries the MIT license of this package, the MIT license of tensoracle, the Apache-2.0 license of the TensorRT headers used to build the plugin, and NVIDIA's notice.
• The wheel metadata no longer carries project URLs, and its CUDA classifier is Environment :: GPU :: NVIDIA CUDA :: 13.
• The README describes installation and use only.
• The Python sources and the native plugin sources are unchanged, and the plugin library in the wheel is identical to the one in v0.10.0. Public names, accepted arguments and outputs are the same as in v0.10.0.