Orinfer 0.2.0 single-opt1 — modified 2026-10-21

This distribution contains a modified Orinfer engine based on upstream commit
a5a2eccfc9a4c8494a012b374728467cb2fa2db9, under the upstream LGPL/GPL terms.
Only crates/orinfer-engine/src/runtime/requests/speculation/control.rs changed:
single-decode-round MTP exploration, one-time cold-cost exclusion, and
request-local acceptance scoring. The runtime enables explore_local.
Model weights and the compiled SM87 execution package are unchanged.

Corresponding complete modified source:
  orinfer-v0.2.0-single-opt1-source.tar.xz
Unmodified upstream source used by the base-image build:
  orinfer-v0.2.0-source.tar.xz
Build and artifact identities:
  orinfer-single-opt1-provenance.json
Patch, native build script, and qualification report:
  https://github.com/manateelazycat/lzc-aipod-qwen38-flash-next/tree/main/orin/single-opt

The runtime image also includes patched-source.tar.xz and
single-opt-provenance.json under /opt/orinfer/source/.
See ORINFER-LICENSE, ORINFER-COPYING, and ORINFER-THIRD-PARTY-NOTICES.

Token statistics: add /metrics and cumulative native completed prompt/output/request counts, shared by JSON and SSE. Preserve execution packages, weights, MTP scheduling and GPU kernels. Corresponding API patch: patches/orinfer-token-statistics.patch.
