LevelZero.NET v2.0.0 - Native Intel GPU Acceleration & Transformer Engine

LevelZero.NET v2.0.0 — Native Intel GPU Acceleration & Transformer Engine

LevelZero.NET v2.0 provides ultra-low latency, zero-allocation C# bindings to the Intel® oneAPI Level Zero runtime, complete with a production-grade Transformer Kernel Suite and an OpenAI-compatible local LLM server.

---

🌟 What's New in v2.0.0

• Transformer Kernel Suite & Multi-Arch Dispatch:

  • Native Intel GPU kernels precompiled into device-native .zebin and portable .spv formats across 11 Intel GPU architectures (Tiger Lake tgllp, DG1, Alchemist acm-g10/g11/g12, Meteor Lake mtl, Arrow Lake arl-h, Lunar Lake lnl-m, Battlemage bmg-g21, Panther Lake ptl-h, Ponte Vecchio pvc).
  • Operators include: rms_norm, rms_norm_quant, rope, gemv_q4_k_f32, gemm_q4_k_f32, softmax, swiglu, silu, and fused multi-head attention.

• l0llm — High-Performance Pure C# LLM Engine & REST Server:

  • Standalone, zero-allocation runtime for LLaMA, Qwen, DeepSeek, and Mistral architectures.
  • Built-in OpenAI-compatible REST server (/v1/chat/completions, SSE streaming, /v1/models, /health).
  • Sharded safetensors (model.safetensors.index.json) support.
  • Achieves 41.3 tok/s decode on consumer Intel Iris Xe GPUs.

• l0check — Level Zero Diagnostic Tool:

  • Comprehensive hardware diagnostic CLI for inspecting GPU topologies, driver versions, memory subsystems, and executing compute kernel test suites.

• NuGet Package:

  • Published to NuGet.org: LevelZero.NET 2.0.0

---

📦 Included Release Assets

• l0llm-win-x64-v2.0.0.zip — Standalone zero-dependency executable (l0llm.exe), interactive chat runner, and OpenAI-compatible server.
• l0check-win-x64-v2.0.0.zip — Standalone Level Zero diagnostic and hardware verification CLI.
• LevelZero.NET.2.0.0.nupkg / .snupkg — Official NuGet package and symbols.

---

🚀 Quick Start (CLI & REST Server)

Interactive Chat

l0llm.exe chat -m path\to\model.safetensors -t path\to\tokenizer.json

OpenAI-Compatible REST Server

l0llm.exe serve -m path\to\model.safetensors -t path\to\tokenizer.json --port 8080

Connect any OpenAI-compatible client, Web UI (e.g. Open-WebUI), or LangChain / ML.Agentic agent to http://localhost:8080/v1.