Release v0.1.0

🎉 v0.1.0 — Hello, vllm Spark2_5 plugin

Every great journey starts with a single commit. Today, Spark2_5 takes its first step into the open.

🚀 What is this?

vllm-spark2_5-plugin brings Spark2_5 model support to vLLM — as an out-of-tree plugin, zero upstream patches required.

Spark2_5 is no toy. It ships with:

• Hybrid sliding-window / full attention — the right attention mechanism for the right context
• Head-wise attention output gating — learned per-head gates before merge
• Grouped-query attention + tensor parallelism — scale across GPUs without breaking a sweat
• Pipeline parallelism — multi-stage inference via IntermediateTensors
• Optimized MLP — fused gate + up projections, GeluAndMul, row-parallel down projection

And a custom XML/KV tool-call parser that turns structured function calls into typed JSON — integer, number, boolean, no stringly-typed nonsense.

✨ What's in v0.1.0

Core

• Model implementation — Spark2_5ForCausalLM, vendored from vLLM commit 81efe78
• Transformers config — Spark2_5Config with full hyperparameter support
• Tool-call parser — registered as spark25 in vLLM's ToolParserManager
• Plugin override — SPARK2_5_PLUGIN_OVERRIDE=1 for explicit vendored-model enforcement

Project Governance

• Apache-2.0 license with proper attribution (NOTICE, THIRD_PARTY_LICENSES.md)
• Code of Conduct & Security Policy aligned with XHToken Community
• CI pipeline: lint (ruff) → test (pytest, Python 3.10–3.12) → build
• Dependabot for pip + GitHub Actions
• Automated PyPI publishing on tag push

📦 Install

uv pip install vllm-spark2_5-plugin

Then serve:

vllm serve /path/to/spark2_5-model \
  --tool-call-parser spark25 \
  --chat-template /path/to/chat_template.jinja

👥 Contributors

• @dongjiang1989 — Plugin architecture, CI/CD, project governance
• @KnightYao — Model vendoring, tool-call parser, early testing and Spark3 → Spark2_5 rename
• @dependabot[bot] — Keeping dependencies fresh

🔮 What's next

This is day one. The foundation is solid — clean plugin architecture, solid test coverage, reproducible builds, automated releases. From here, we follow vLLM upstream, adapt to its evolution, and push Spark2_5 performance further.

PyPI Package: https://pypi.org/project/vllm-spark2-5-plugin
Full Changelog: https://github.com/XHToken/Spark-plugin/commits/v0.1.0