cuda-13

MSVC+CUDA llama.cpp fork for Windows: KDA/GDN, long-context KV placement, TurboQuant/TCQ when measured, speculative draft. Lab defaults — TriAttention opt-in only, not recommended.

Qwen3.8-Flash-Next (NVIDIA NVFP4 weights) served as W4A16 on 2x DGX Spark GB10 with vLLM, TP=2. Pinned engine build, 4 start-time patches, measured results, and the levers that were tried and rejected.

Production-grade SGLang inference on NVIDIA DGX Spark / GB10 (sm_121) — sm_121a-native sgl-kernel, model-agnostic launch recipes, wedge handling.

Pre-built PyTorch wheels and build scripts for NVIDIA DGX Spark (GB10, sm_121, Blackwell, CUDA 13.0, ARM64)

Measured sm_110 / sm_110a facts for NVIDIA Jetson AGX Thor: tensor-core capability matrix (2:4 sparsity works on plain sm_110; tcgen05 + NVFP4 block-scale is sm_110a-only), CUDA 11/12->13 migration breaks, CPU ISA, and the bandwidth roofline. With probes.

NATTEN 0.21.6 Windows wheel for Python 3.13, PyTorch 2.10, CUDA 13.0, NVIDIA Blackwell sm_120 / RTX 5090; Pixal3D-ComfyUI compatible.

Build paddlepaddle-gpu from source for NVIDIA DGX Spark and GB10 (Grace Blackwell, aarch64, CUDA 13, sm_121). Prebuilt wheel plus the exact patches upstream is missing.

Build the GPU inference stack from source, repeatably: for Blackwell sm_120 on CUDA 13.x and Python 3.14. Machine-readable build state, generated patches, and an abductive-triage skill for build failures.

Containerized vLLM 0.28.0 + Qwen3-1.7B thinking model on NVIDIA DGX Spark (GB10, sm_121, 128GB unified memory). Reasoning parser qwen3 splits the think block from the user answer (message.reasoning vs message.content). Includes scripts to launch, tail logs, and run a think/content split test, plus an architecture manual and HANDOFF notes.

Hunyuan3D-2 fork — image→textured 3D→sliced STL + part segmentation. RTX 50-series (Blackwell/sm_120), CUDA 13.0, Python 3.12, PyTorch 2.11+cu130.

Windows NVIDIA-only Triton 3.7.0 build pipeline for RTX 5090 / Blackwell sm_120a, with FP8 tl.dot validation and peak benchmark results.

Community FlashAttention fork for CUDA 13 and NVIDIA Blackwell SM120 inference workloads

Automated Docker build and serving stack for vLLM with native NVFP4 KV cache quantization on NVIDIA Blackwell (RTX 5090 / SM120)

Accelerate Kimi Delta Attention computations with high-performance CUTLASS kernels designed for NVIDIA SM90 architectures and beyond.