prismml

HIP/ROCm fork optimized for AMD RDNA2 (gfx1030) with PrismML Q1_0_G128 1-bit quant support, RotorQuant, TurboQuant, EAGLE3 and P-EAGLE speculative decoding, and full Wave32 kernel optimizations.

Precompiled PrismML/llama.cpp backend for LM Studio on Windows. Run PrismML Bonsai 1-bit/ternary GGUF models natively in LM Studio with NVIDIA CUDA. No compiling required.

AMD ROCm (gfx1030) inference fork with RotorQuant/TurboQuant KV compression, PHANTOM-X zero-copy draft speculation, EAGLE3 speculative decoding, 12 RDNA2 crash fixes, and PrismML Bonsai Q1_0_G128 1-bit GGUF support.

Ternary Bonsai 2: New 0.4GB Vision AI With 310 Tokens/Sec Speed! - 100% private visual document scanner and PII redactor running locally.

HTTP proxy serving PrismML Bonsai 1.7B (Q1_0 GGUF) through an Ollama-compatible API via llama-server.

Windows x64 installer for Bonsai 2 / PrismML GGUF models in LM Studio. Isolates PrismML server binaries, preserves native libraries, verifies downloads, and backs up runtime manifests.

Navigable Degeneracy in the Roots of 1-Bit Language Models

One-click Windows setup & launcher for PrismML Bonsai-demo — pick a model, everything else is automated.

GGUFly — Interactive TUI launcher and model manager for GGUF models with llama.cpp, PrismML/Bonsai, and Mirai S runtimes. Built on Omarchy / Arch Linux.

Restart-proof KV cache for llama.cpp: 40K-token TTFT 155s→6s after reboot. llama-server 重启/关机后自动恢复 prompt 前缀缓存，不重算 prefill。

Ternary Bonsai 2 9B on a GTX 1080 (sm_61): +51% decode with bit-identical output (PrismML llama.cpp fork, PQ2_0 + MTP)

Local chat app for running Bonsai GGUF models with streaming responses, built-in Prism runtime management, model switching, runtime diagnostics, and saved conversation history.

Local deployment of Ternary Bonsai 2 27B (PrismML 1.75-bit ternary quantized Qwen3.8 27B) on an RTX 4060 laptop — setup, benchmarks and notes

A local, air-gapped multi-agent coding harness: sequential agent roles on local models, tools in a Podman sandbox, and nothing red ever committed.

Ternary-Bonsai-2-27B (PTQ1_0 QAT) on 8GB - PrismML b10709 A/B (+15% PP/+24% TG), quality gate, production config

Byte-level reverse engineering of PrismML Ternary Bonsai 2 27B: 1.58-bit quantization, 42/128 zero mask, Hadamard rotation, GGUF formats, and fact-checked benchmarks.

Reproducible local Bonsai model server for Hermes: auto-detects your hardware, fetches official PrismML prebuilt binaries and models, and wires up the OpenAI-compatible endpoint.

Terminal workspace for chatting with Bonsai AI models locally using the PrismML custom runtime.