rtx-4090

run a 27b local ai on your gaming gpu, rtx 3060 and up, radeon rx 6600 and up in beta, 8gb or more, linux and windows, one line to install

SNDR Core Engine (Genesis) — vLLM runtime patch-overlay for Qwen3.6 + Gemma4 on consumer NVIDIA (Ampere sm_86, 2× A5000/3090). Qwen3.6-35B-A3B FP8 ~240 tok/s, 27B-int4 hybrid GDN+Mamba, Gemma4 26B/31B AWQ, 256K ctx. 321 patches: TurboQuant k8v4 KV, MTP/DFlash spec-decode, FULL cudagraph, hybrid GDN. vLLM pin dev424 + Control Center GUI.

Durable local inference for Oh My Pi: NInfer + Qwen3.8 27B on one RTX 3090/4090/5090, with restart-resumable OpenAI Responses state.

异环（Neverness To Everness / Ananta）光线追踪一键部署面板，基于 OptiScaler winmm 方案，默认推荐 RTX 5090，并支持本机/RTX 4090/RTX 5080M 配置、备份、恢复和本地 WebUI。

Unlocks RT Options in NTE Global PC for Low-Mid RTX Graphics Cards

Reproducible llama.cpp configs + per-category quality benches for Qwen3.6-27B on a single RTX 4090. Winners, dead ends, and the silent-corruption bug.

Qwen3.6 vLLM Toolkit — Launcher + Templates, Optimized for 2×24 GB. vLLM launcher with an optimized hybrid chat template drawing from the best community fixes for Qwen 3.6 27B.

面向 8×RTX 4090（SM89）的 DeepSeek-V4-Flash-0731 MXFP4 SGLang 实验分支

Quality-first local MiniMax H3 video-series generation and loopback API for dual RTX 4090 workstations, with native audio, references, P8/P9 continuity, and preserved artifacts.

4K video super-resolution on a 24 GB RTX 4090 — a free, local ComfyUI FlashVSR node (VIDEO in, VIDEO out) plus a Triton/FP8 operator pack at ~1.3×. No CUDA build, every number backed by benchmark JSONs.

Hardware-adaptive local LLM & cloud cascading gateway. Sub-5ms intelligent routing, 3-token lookahead failover, asymmetric verification, 1-click IDE config, and MCP for $0 token cost.

QSA HiSparse as an SGLang plugin: 256K CPU KV offload and text/image host prefix reuse for Qwen Sparse Attention on unmodified SGLang

🚀 Advanced UniPC sampler surgical control node for ComfyUI. Exposing internal Predictor-Corrector mathematics to manage fine detail (D-tail) and noise trajectory. Precision sampling control for high-end AI workstations.

GPU JPEG2000 (J2K) benchmark: fvJPEG2000 vs nvJPEG2000 on an NVIDIA RTX 4090. Reproducible measurement procedure with scripts and logs.

A direct memory-to-RAM mapping protocol for PyTorch/CUDA. Bypasses the swap file/SSD page file for zero-latency AI computations. Optimized for RTX 4090/5090 and up to 128GB of RAM and higher.

Memory-aware multimodal video restoration agent with Qwen2.5-VL, auditable routing, OOD abstention, multi-tool restoration, and closed-loop quality control on a single RTX 4090.

Dual NVENC batch video compressor — leverages RTX 4090's twin NVENC chips for parallel AV1/HEVC encoding

Benchmark open-weight LLMs on vLLM in Docker under one fixed protocol: decode at 0/4k/16k, cold TTFT, 4-way concurrency, power and VRAM, three runs with spread. Standard-library Python, charts and a results explorer.

Production-grade SFT + ORPO fine-tuning scripts for small reasoning & instruction models. Built with Unsloth + TRL, stabilized for single-GPU (RTX 4090) training with long thinking tokens. More models coming.

Stock Oh My Pi on a stock, local Strata inference server: one Windows 11 + RTX host, every component hash-pinned, loopback behind an API key, and real-host acceptance evidence for every claim. Nothing forked, no proxy.