Skip to content

Version 1.1.2.0

Latest

Choose a tag to compare

@Integral2u Integral2u released this 20 Nov 08:20

The first GitHub release since 1.0.9.0. Versions 1.1.0 and 1.1.1 were written up in the changelog and published to NuGet but never given a release here, so this one covers all three.

Models and architectures

  • Multi-head Latent Attention (DeepSeek V2/V3, Moonlight) — MLA keeps the
    compressed KV path end to end, with the DeepSeek V3 routing knobs
    (expert_gating_func), a dense leading block in a routed file, and shared
    experts counted from the file rather than assumed.
  • YaRN rotary scaling — DeepSeek-V2/V3 RoPE now blends interpolated and
    extrapolated angles across the correction band and folds the YaRN attention
    factor into the tables, so long-context position stops being guesswork.
  • MXFP4 (GGML type 39) — decode plus vec-dot and matmul kernels on x86
    (scalar, AVX2, FMA) and Arm64, and kbench mxfp4.
  • More quant formats — IQ4_XS decode, and the grid formats IQ2_XXS, IQ2_XS,
    IQ3_XXS, IQ2_S and IQ3_S wired into layer dispatch alongside IQ1_S, IQ1_M and
    TQ1_0.
  • LFM3 tri-hybrid, LFM3-Audio (encoder, ASR including long recordings,
    native TTS) and LFM3-VL vision through the mmproj path.
  • gemma3n, qwen35moe and llama4 resolve to their real presets.

Speed

  • Arm64 NEON vec-dot and matmul for Q8_0, Q4_K and Q6_K, with an Arm64 CI
    runner and HardwareTier.Arm64 behind it.
  • Fusion and dispatch fixes: Q4_0, Q4_1, IQ4_NL and Q8_K now actually use their
    FMA/SSE kernels in a real model, IQ1_S/IQ1_M/TQ1_0 no longer decode through
    the F32 kernel, and a linear layer's kernel suffix follows the mapping's tier
    instead of the host.

Serving and tools

  • Server admission control — a concurrency cap and a bounded queue.
  • Audio primitives — arbitrary-size FFT (Bluestein), RIFF/WAVE read and write.
  • Quantization and analysis — GgufQuantizer re-encodes a GGUF without
    loading it, plus layer-importance, latent-projection, activation-map,
    skip-layer and quantization-comparison analyzers.
  • Prompt-lookup drafting — opt-in GenerationConfig.PromptLookupDraft.
  • KV snapshots record the prompt tokens they hold.
  • Delegation keeps the parent cache; one-shot sub-agents no longer start cold.
  • Training — record-aware ITrainingBatchSource, a TrainLoop overload and
    an SFT path, plus LayerNorm and dense GELU FFN on the GPU inference engine.

Fixed

  • An MLA file no longer loads as plain attention.
  • An ungated shared expert was being halved.
  • YaRN rope tables lost 27% of every rotated dimension.
  • PermissionPathPolicy case sensitivity follows the platform.
  • Model loads retry a Unix sharing conflict.
  • SummarizingCompactor keeps system prompts and the current turn.

Installers

  • SharpMind CLI 1.1.2 (SharpMind.CLI.Setup.msi) — framework-dependent;
    needs the .NET 10 and ASP.NET Core runtimes.
  • SharpMind Console 1.1.2 (SharpMind.Console.Setup.zip) — the TUI,
    packaged as an MSI inside a zip.

Both carry a new ProductCode/PackageCode pair on the same UpgradeCode, so they
upgrade a 1.0.9 install rather than sitting beside it.

The NuGet packages (SharpMind.Core, Data, Model, Tokenization,
Inference, Training) are on nuget.org at 1.1.2.

Full changelog: https://github.com/Integral2u/SharpMind/blob/master/CHANGELOG.md