sgBLAS is a project of Abhinav Gorrepati. Its pre-publication development used AI coding assistants, including OpenAI Codex, for activities that may have included planning, implementation drafts, tests, documentation, review, and debugging. Abhinav selected, reviewed, integrated, and takes responsibility for the material released in this repository.
This disclosure does not claim that every line was independently handwritten. AI-generated suggestions are treated as untrusted drafts: they do not establish originality, correctness, or license compatibility, and they must receive the same human review as any other contribution.
The public Git history begins with the first reviewed release snapshot. Earlier intermediate commits are not available as an authorship record. The phrase "from scratch" means that no third-party SGEMM implementation is vendored or linked into sgBLAS itself; it does not mean that the project was developed without documentation, architectural references, standard APIs, or AI assistance.
The project uses NVIDIA's public CUDA, cuBLAS, PTX, and CUTLASS documentation as
technical references. CUTLASS is reference-only and no CUTLASS code is vendored.
External code may be incorporated only when its provenance and license are
known, redistribution is compatible with this project, and all required notices
are preserved in THIRD_PARTY_NOTICES.md or another appropriate notice file.
Contributors are responsible for identifying any copied, adapted, translated, or generated material in their changes. When provenance is uncertain, the material must not be merged until ownership and licensing are resolved.
A correctness, sanitizer, or performance claim is considered reproducible release evidence only when its artifact bundle records all of the following:
- the exact Git commit SHA from a clean working tree;
- a deterministic source-tree digest and hashes of the tested binaries;
- the container image digest or complete native toolchain and operating-system versions;
- GPU model, compute capability, driver, CUDA runtime, CUDA compiler, and cuBLAS versions, with sensitive stable identifiers removed from public artifacts;
- the exact commands, configuration flags, seed, warmup count, repeat count, cache policy, and relevant environment variables;
- raw correctness, sanitizer, and benchmark logs plus a checksum manifest; and
- a generated summary whose values can be traced back to those raw logs.
Any source, build configuration, toolchain, or measurement-protocol change breaks that binding. Affected claims must be rerun or explicitly labeled as historical evidence for a different snapshot.
Benchmark and sanitizer results created before the first public commit were measured from an uncommitted development snapshot. They may be useful as historical development evidence, but they are not release-reproducible evidence for a later commit until the full validation campaign is rerun and bound as described above.