Skip to content

UPSTREAM.md

Latest commit

 

History

History
38 lines (29 loc) · 1.9 KB

UPSTREAM.md

File metadata and controls

38 lines (29 loc) · 1.9 KB
Raw

Upstream and local change record

NInferEZ Engine is a thin Windows distribution fork of iamwavecut/ninfer-all. It is not affiliated with or endorsed by NInfer or the upstream maintainers.

Current base

  • Upstream repository: https://github.com/iamwavecut/ninfer-all.git
  • Base commit: 91576f32ee6147c31d5ead49393cce049530b6e1
  • Base commit date: 2026-09-27
  • First NInferEZ Engine release line: 0.1.0-preview.1

The repository keeps the upstream Git commit as its base and uses an upstream remote. Upstream authors, contributor records and provenance documentation remain in the tree.

Intentional local changes

  1. Replace the GCC-only __builtin_ctz use in a CUDA host path with the C++20 std::countr_zero equivalent so MSVC builds the same operation.
  2. Generate built-in device-profile JSON from bounded string chunks to avoid MSVC C2026 without changing the JSON payload.
  3. Split the serving option parser into two sequential groups to avoid MSVC C1061 without changing accepted flags or defaults.
  4. Accept third-party GSQ-RCO GGUF exports that retain ssm_alpha.weight in F32 and convert only that small direct projection to BF16. Quantized GGUF block matrices remain byte-for-byte copied.
  5. Add a target-aware Windows build/release pipeline for sm86, sm89 and sm120a.
  6. Add stable JSON identity/capability commands and a host-only .ninfer inspector for NInferEZ and other clients.

None of these changes inserts a proxy into the inference path. Model execution, HTTP serving and CUDA kernels continue to run directly in the NInfer engine.

Updating from upstream

An upstream update must be merged as history, not copied over the tree. Reapply only the local changes still needed, then build every target. Runtime qualification is recorded separately per GPU family: a successful compile is Preview evidence, not Stable hardware qualification.