GPU-Initiated Communication: Dissecting Down to the Bone
GPU 发起的通信允许 GPU 线程直接将 RDMA 操作发布到网卡。它是 NVSHMEM、NCCL GIN 和 DeepEP 的基础,这些技术支撑着混合专家(MoE)模型中细粒度、低延迟的关键通信需求,但其性能特征和优化方法至今鲜有深入研究。
// 主题信号
训练框架、推理引擎、向量数据库、GPU 调度、可观测性等支撑技术。
GPU 发起的通信允许 GPU 线程直接将 RDMA 操作发布到网卡。它是 NVSHMEM、NCCL GIN 和 DeepEP 的基础,这些技术支撑着混合专家(MoE)模型中细粒度、低延迟的关键通信需求,但其性能特征和优化方法至今鲜有深入研究。
Add "Running LLMs Locally" sub-section under Guides & References / AI: why and when to run on premise, hardware sizing (VRAM, KV cache, quantization), Ollama, llama.cpp, vLLM / SGLang, desktop runners (LM Studio, Jan, Docker Model Runner, LocalAI, MLX-LM), NVIDIA NIM, open-weight model families and licences, LLM gateways (LiteLLM), inference optimisation, and integrating local models with GitHub Copilot (BYOK), Claude Code (LLM gateway / Anthropic-compatible endpoints) and other coding assistants, with MathJax, bibliography and a one-page cheat sheet PDF
Add "AI" section under Guides & References and its first sub-section "LLM Foundations & Prompting": AI landing page, how LLMs work (tokens, transformers, training, sampling, context, reasoning, multimodality, hallucination), prompt and context engineering, structured output, choosing models (benchmarks, cost/latency, open vs. closed), with MathJax, bibliography and a one-page cheat sheet PDF
Cargo-atlas – A compiler-accurate Rust code map for AI coding agents
如果它还能在每次更改模型时自动配置/重新配置 Hermes 中的自定义端点,那就太棒了。我针对该面板(APCA 对比度、布局节奏、按钮居中)提出了设计改动的 PR:https://github.com/0xSero/omarchy/pull/2
现代 GPU 内核日益对指令供给路径造成压力,而固定指令容器可能留下大量体积冗余。本文提出了一种自动编码合成框架,将指令布局视为经过验证的指令之上的约束槽位分配问题