Skip to content

Releases: MinsungMJ/Local-Inference-System

Release list

LIS v0.1.1 — CI and Documentation Update

LIS v0.1.1 — CI and Documentation Update Latest
Latest

Choose a tag to compare

View all tags
@MinsungMJ MinsungMJ released this 08 Jun 06:06
v0.1.1
2510afa

LIS v0.1.1 adds automated validation and clarifies the documented Qwen3 prompt boundary. Runtime and model-support scope remain intentionally conservative.

What changed

  • Added GitHub Actions CI on Ubuntu 24.04
  • CI validates the C build, C test suites, and Python LIS Inspect tests
  • Added a CI status badge to the README
  • Documented that the Qwen3 path currently receives raw tokenizer text
  • Clarified that LIS does not apply model-specific chat templates or expose thinking-mode controls
  • Clarified that reasoning-oriented models may produce extended explanatory output under the raw-prompt and greedy-decoding path
  • Generalized local maintainer-note ignore patterns

Scope remains unchanged

  • CPU-only local inference
  • Greedy decoding only
  • Documented Llama-family plain-RoPE subset
  • Narrow Qwen3 Dense BF16 path
  • No GPU execution, serving, distributed inference, continuous batching, sampling framework, or broad Qwen-family claim
    See the tagged README, v1 Support Matrix, and Qwen3 Dense Scope.

License

Apache License 2.0

Assets 2
Loading

LIS v0.1.0 — Initial Public Release

LIS v0.1.0 — Initial Public Release

Choose a tag to compare

View all tags
@MinsungMJ MinsungMJ released this 08 Jun 06:05
v0.1.0
d567c59

LIS v0.1.0 is the first public release of the Local Inference System.
LIS is a CPU-only inference runtime for causal decoder-only models, designed around correctness, inspectability, reproducible artifacts, and transparent diagnostics.

What’s included

  • CPU reference execution for the documented Llama-family plain-RoPE subset
  • A narrow Qwen3 Dense BF16 merged-safetensors path
  • Deterministic greedy decoding
  • Hugging Face tokenizer JSON subset support
  • Opt-in execution reports, traces, diagnostics, and performance breakdowns
  • LIS Inspect for reviewing run_report JSON and optional perf stderr logs
  • Model-free make build and make test

Known limitations

  • Greedy decoding only
  • No sampling framework, GPU execution, serving, distributed inference, or continuous batching
  • No model-specific chat-template or Jinja execution
  • No thinking-mode control
  • The Qwen3 path is intentionally narrow and does not imply broad Qwen-family support
    See the tagged README and v1 Support Matrix for details.

License

Apache License 2.0

Loading