AI infrastructure
End-to-end technical blueprint for enterprise AI infrastructure & inference engineering: GPU memory mechanics, interconnect topologies, quantization, PagedAttention, and model router selection frameworks.
Patches enabling NVLink symmetric-memory all-reduce and related fast paths on Ampere (sm_80/sm_86) for vLLM and FlashInfer
deploy & experiments LLM at production scale
Safety gate for AI agents that change infrastructure — asks first, signs every decision into a tamper-evident ledger, probes the network to catch drift. Zerops Challenge 2026.
Run MiniMax H3 at native fp16 on pre-bf16 GPUs (V100/Volta) — exact-math fix for the black-frame overflows, ~11× faster than the fp32 fallback.
本源计算架构 · An open architecture for persistent AI computers. 模型可换，状态不丢，动作可迁移，经验会复利。
Generic inference hook that runs on Google Cloud Run. Compatible with Claude Enterprise Inference Hooks.
A streaming Mixture-of-Experts (MoE) inference engine for the GPT-OSS models (20B and 120B), written in pure C, that runs models larger than your RAM on ordinary consumer hardware.
Session identity, independent mode counters, and emit-time headers for multi-turn assistant conversations. Instrumentation contract by Andrew Bradbury. Not an agent runtime.
Atomic Redis spend cap for LLM APIs (OpenAI, Gemini, …) — one Lua-atomic INCR+PEXPIRE so a bug or abuse can't blow up your bill.
This port of RAM coffers adds Kimi K3 inference on PlayStation 3 clusters - about 100 thousand nodes needed to run it
FAAAH (Filesystem As An AI Handler) - Reuse your AI Agent subscription via text files.
Two LLMs in one weight stream: polar dual-model packing, 1.78x compression, near-lossless quality
MiniMax H3 视频生成模型压力测试方法论 — 六模式 24h 持续压测框架，普适于任意自部署视频生成模型（文生视频/图生视频/多模态参考生成）
Project Kalos: A Biological-Resonant C/CUDA Neuromorphic Engine Suite & Empirical Whitepapers v1.0
setup guide for deepseek-flash-0731 on two dgx-spark from scratch
Run DeepSeek-V4-Flash at full 1M context AND two MiniMax H3 video instances on the same two DGX Sparks. Benchmarked C1-C6.
Open recipes, engine patches, and benchmark harnesses for LLM inference on Intel Arc Pro B60/B70 (Battlemage, Xe2). MoE 35B at 126 t/s decode / 7.5K t/s prefill single-stream — vLLM XPU MTP unlocked.