搜索结果：DeepSeek

PLENA is a hardware-software co-designed system for LLM agentic inference that addresses bandwidth and capacity memory walls. It features a flattened systolic-array architecture, asymmetric quantization, and FlashAttention support, achieving up to 2.23x and 4.70x throughput improvements over A100 GPU and TPU v6e, respectively, and 4.04x better energy efficiency than A100. 原文翻译： PLENA是一个硬件-软件协同设计的系统，针对LLM代理推理，解决带宽和容量内存墙问题。它采用扁平化脉动阵列架构、非对称量化和FlashAttention支持，相比A100 GPU和TPU v6e，吞吐量分别提升2.23倍和4.70倍，能效比A100提升4.04倍。

This article introduces a comprehensive GEO (Generative Engine Optimization) methodology, focusing on expert Yu Lei's 'Two Cores + Four Drivers' system. It evaluates multiple GEO approaches, provides

Chip Huyen's blog covers key topics in AI engineering, including common pitfalls in building generative AI applications, agents, generative AI platforms, and LLM research challenges. The content is highly relevant for technical professionals building production AI systems. 原文翻译：Chip Huyen的博客涵盖了AI工程的关键主题，包括构建生成式AI应用的常见陷阱、智能体、生成式AI平台以及LLM研究挑战。内容对构建生产级AI系统的技术专业人士极具参考价值。

llm-d is a high-performance distributed inference serving stack optimized for production deployments on Kubernetes. It achieves SOTA inference performance across various accelerators by integrating vLLM, Kubernetes Gateway API, and advanced orchestration techniques such as disaggregated serving, prefix-cache aware routing, and tiered KV caching. The v0.5 release demonstrates up to 50k output tok/s on a 16×16 B200 topology. 原文翻译： llm-d是一个针对Kubernetes生产部署优化的高性能分布式推理服务栈。它通过集成vLLM、Kubernetes Gateway API以及分离式推理、前缀缓存感知路由、分层KV缓存等高级编排技术，在各种加速器上实现SOTA推理性能。v0.5版本在16×16 B200拓扑上展示了高达50k输出tok/s的性能。

This article presents a local hybrid RAG pattern combining lexical retrieval, ONNX-based semantic embeddings, and Foundry Local chat model for offline AI assistants. It covers architecture, implementation, and best practices for graceful degradation when semantic path fails. 原文翻译：本文介绍了一种本地混合RAG模式，结合词法检索、基于ONNX的语义嵌入和Foundry Local聊天模型，用于离线AI助手。涵盖架构、实现和最佳实践，确保语义路径不可用时优雅降级。

Ssebowa is an open-source Python library offering generative AI models for text, image, and video generation, including LLM, VLLM, image generation, and video generation. It supports fine-tuning with

RAG-Anything is a lightweight RAG system based on LightRAG, designed for multimodal document processing (PDF, images, tables, formulas, etc.). It provides end-to-end parsing, multimodal understanding,

RAG-Anything is an all-in-one multimodal RAG system that processes documents containing text, images, tables, and formulas. It features end-to-end processing pipelines, knowledge graph indexing, and c

This paper introduces BlockRank, a method that exploits attention sparsity in LLMs for in-context ranking, reducing complexity from quadratic to linear and enabling efficient retrieval of up to 500 documents within a second. 原文翻译：本文提出BlockRank，利用LLM注意力稀疏性进行上下文排序，将复杂度从二次降至线性，实现秒级检索500个文档。

This article explains the critical importance of structured outputs in LLM workflows, detailing how to implement them from scratch using JSON and Pydantic, and through the Gemini SDK, to build reliabl