clee.sh / qxmx
A single-model (Bonsai-27B) inference engine for Intel Arc GPUs
Star
Fork 1
Watch
qxmx third_party
6 files
httplib.h
README.md
unicode-data.cpp
unicode-data.h
unicode.cpp
unicode.h
README.md

third_party — vendored MIT-licensed code #

unicode.{cpp,h}, unicode-data.{cpp,h} #

Copied verbatim from llama.cpp (https://github.com/ggml-org/llama.cpp), MIT-licensed, Copyright (c) 2023-2026 The ggml authors.

Provides Unicode code-point classification (\p{L} \p{N} \p{M} \s etc.) and the hand-rolled pre-tokenizer regex splitters (unicode_regex_split, including the qwen35 variant this model needs). The data tables are machine-generated; re-implementing them in C would risk transcription errors, so they are vendored unchanged and linked as C++.

See ../../LICENSE-third-party.txt for the full MIT notice.

Used by tokenizer.cpp for BPE pre-tokenization.

httplib.h #

Copied verbatim from cpp-httplib v0.18.3 (https://github.com/yhirose/cpp-httplib), MIT-licensed, Copyright (c) 2024 Yuji Hirose. A single-header HTTP/HTTPS server+client; we use only the server side for the qxmx_serve OpenAI-compatible endpoint (M3). No OpenSSL, zlib, or brotli support is enabled.