Cross-platform installer for Triton and SageAttention on ComfyUI. Simplifies GPU-accelerated inference setup for Windows users with automated dependency management and RTX 5090 support.

Tiny local text-to-24x24 pixel art model, trained on roughly 200K samples in 30 minutes on an RTX 5090.

From-scratch C++/CUDA inference engine for the NVIDIA RTX 5090 (sm_120a) — the best single-GPU backend for agentic AI: tool calling, long-context loops, reasoning and concurrent sub-agents on top of the fastest single-stream decode on the 5090 (beats llama.cpp, at-or-ahead of vLLM on NVFP4). 100% written by Claude Code.

RTX 5090 & RTX 5060 Docker container with PyTorch + TensorFlow. First fully-tested Blackwell GPU support for ML/AI. CUDA 12.8, Python 3.11, Ubuntu 24.04. Works with RTX 50-series (5090/5080/5070/5060) and RTX 40-series.

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。

Pixal3D ComfyUI integration for Windows (RTX 30/40/50) — single image to textured PBR mesh in 3-5 min

NVFP4 inference on Blackwell GeForce (RTX 5090/5080/5070 Ti/RTX PRO 6000) — SM120 patches for vLLM + FlashInfer + CUTLASS. 175 tok/s on Qwen3.6-35B MoE.

The old advice to avoid Windows for local LLMs used to be right. It isn't anymore. The fixes for display-GPU desktop crashes and WSL memory limits, from people who run a 5090 daily.

Research: vGPU unlock on consumer NVIDIA RTX 5090 (Blackwell/GB202). 19 binary patches, full CPU-side pipeline working, GSP firmware blocked by fused-off VF PRIV registers.

Fastest MoE/LLM inference runtime for consumer and edge Blackwell GPUs. SN74 on Gittensor.

Dashboard for AI Studio, Open Source Continuous Inference | Deepseek-R1, Qwen2.5, Llama3.1 | 4xRTX-5090 inside PRU2500, 2xH100 inside PRU2500, 8xMI210 in SuperMicro

Windows prebuilt of llama.cpp combining Multi-Token Prediction (MTP) + TurboQuant KV cache compression + native sm_120 (Blackwell consumer GPU, FP4 tensor cores). For RTX 5060 Ti / 5070 / 5080 / 5090.

LLM benchmarking, GPU workload orchestration backend server | Deepseek-R1, Qwen2.5, Llama3.1 | 4xRTX-5090 inside PRU2500, 2xH100 inside PRU2500, 8xMI210 in SuperMicro

Fork of NVIDIA MinkowskiEngine modernized for CUDA 12.8+/Blackwell (RTX 50-series), NumPy 2.0, PyTorch 2.x — prebuilt wheels, fp16/bf16 autocast, fused gather/scatter. Best-effort, no support commitment.

A high-performance local AI pipeline for restoring VHS audio, transcribing with Whisper, and translating subtitles using NLLB-200.

Fish Audio OpenAudio S2-Pro on vLLM-Omni. low-latency ~100ms TTFA, OpenAI-compatible, runs on NVIDIA Blackwell (RTX 5090 / RTX PRO 6000). Self-hosted streaming TTS & voice cloning.

CastelOS public artifacts — principles, architecture insights, and build-in-public content

Optimized CSM-1B TTS pipeline for RTX 5090 (Blackwell sm_120). CUDA graph replay via patched HF Transformers. ~0.46x RTF. Topics (tags): csm text-to-speech rtx-5090 blackwell cuda-graphs torch-compile sesame streaming pytorch

Production-grade Traditional Chinese / Taiwan Mandarin speech-to-text. Qwen3-ASR + MediaTek Breeze-ASR-25, hot-word injection, LLM polish, speaker diarization. RTF up to 1554x on RTX 5090, 56 TDD tests.

Drop-in exact bf16 flash-attention for CUDA with a deterministic backward, tuned for Blackwell (sm_120 / RTX 5090).