Skip to content

v0.1.1

Choose a tag to compare

View all tags
@farhadrgh farhadrgh released this 07 Jul 15:15
· 9 commits to main since this release
v0.1.1
5129112
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

nvSubquadratic 0.1.1

Dependency hygiene release — no behavior or public-API changes.

Changed

  • Accelerated CUDA kernels are now an opt-in [cuda] extra. subquadratic-ops-torch-cu12 is a source-only sdist that needs nvcc to build; as a core dependency it made pip install nvsubquadratic fail in environments without a CUDA toolchain (e.g. a downstream project's CPU CI). It's now pip install 'nvsubquadratic[cuda]'. The kernel is imported lazily; the default fft_backend="torch_fft" path is portable and needs no CUDA kernel. Selecting fft_backend="subq_ops" without it raises a clear ImportError.
  • megatron-core → [distributed], timm → [baselines]. Neither is needed to import or run the operators. init_parallel_state() / ConvNeXt-baseline DropPath raise clear ImportErrors pointing to the right extra when absent.
  • Dropped protobuf and huggingface_hub from the explicit dependency list (not imported anywhere; arrive transitively via wandb / datasets).
  • Extras are now [cuda], [quack], [dali], [distributed], [baselines], and [all] (union of all five).
  • Requires-Python lowered to >=3.10 (was >=3.11); CI byte-compiles on 3.10/3.11/3.12. The wheel is pure-Python (py3-none-any).

Install: pip install nvsubquadratic (GPU-first, batteries-included) · full PyPI listing: https://pypi.org/project/nvsubquadratic/0.1.1/

Assets 2
Loading