This article walks programmers through 60 core AI concepts—from the basics of large language models and tokens to advanced topics like prompt engineering, retrieval‑augmented generation, fine‑tuning, and inference optimization—organized into progressive skill levels and illustrated with concrete examples and code snippets.
The paper introduces ECHO, an elastic speculative decoding framework that treats token verification as a global budget‑scheduling problem, uses sparse confidence gating and a two‑level priority scheduler, and demonstrates up to 14.4% throughput gains for high‑concurrency LLM serving.
The article analyzes how the PRISM framework redesigns test‑time scaling for discrete diffusion language models by replacing costly Best‑of‑N sampling with a three‑stage hierarchical search, local branching via partial remasking, and self‑verified feedback, achieving large accuracy gains on math and code benchmarks while cutting inference compute by up to four‑fold.
The article explains why large‑language‑model inference is bottlenecked by memory‑bandwidth, then details Google’s Gemma 4 MTP technique—using a small draft model with speculative decoding and shared KV‑Cache—to parallelize token prediction, achieving up to three‑fold speed gains without any loss in output quality, and provides step‑by‑step local deployment instructions.
RouteMoA moves model selection ahead of inference by using a lightweight scorer to predict each model's suitability from the query, dramatically cutting computation cost and latency while preserving or improving accuracy, as demonstrated on a 15‑model pool with up to 90% cost reduction and 64% latency reduction.
The article introduces the TRS (Thinking with Reasoning Skills) framework, which distills historical LLM reasoning traces into reusable skill cards, enabling offline skill‑base construction and online retrieval that dramatically reduces token consumption (6‑59%) and often improves accuracy on math and coding tasks.
The paper introduces Squeeze Evolve, a validator‑free multi‑model evolutionary framework that orchestrates diverse large language models to break the performance ceiling of any single model, delivering up to 23‑point accuracy improvements and 1.4‑3.3× cost reductions across math, vision, and scientific benchmarks.
On April 20, the open‑source Kimi K2.6 model debuted with industry‑leading code generation, long‑range task execution and a 300‑agent cluster, while Huawei Cloud’s KV‑Cache‑Aware scheduling cuts TTFT by 10% and enables free, one‑click API access for developers.
The article analyzes why the massive KVCache bandwidth required by heterogeneous pre‑fill/ decode (PD) separation cannot be solved at the system level, proposes a Prefill‑as‑a‑Service architecture that leverages linear‑attention models to cut KVCache generation, and validates the design with a 1‑trillion‑parameter Kimi Linear deployment that achieves 54% higher throughput and 64% lower P90 TTFT across a 100 Gbps inter‑datacenter link.
This open‑source tutorial breaks down large language model internals into 11 detailed topics—covering BPE tokenization, attention mathematics, backpropagation, transformer architecture, KV‑Cache, Paged and Flash Attention, and frontier techniques—each with numeric derivations and Python code, making it ideal for developers and interview preparation.
The article explains how KV cache eliminates redundant computations in autoregressive LLM generation, detailing the attention mechanism, the O(n²) waste of recomputing K and V, the cache‑based solution, its impact on time‑to‑first‑token, and the memory‑vs‑speed trade‑off.
The article details how speculative decoding—using a fast small model to draft tokens and a large model to verify them—overcomes the memory‑bandwidth bottleneck of autoregressive inference, introduces SSD’s self‑draft and tree‑verification stages, presents real‑world benchmark gains, and shows how to enable it in vLLM.
The newly released GigaWorld-Policy world‑action model replaces traditional video‑prediction‑heavy WAM designs with an action‑centered architecture, achieving a ten‑fold inference speedup, ten‑fold training efficiency gain, and a 30% increase in real‑robot task success rate while reducing memory usage compared with Motus and Cosmos‑Policy.
Google’s TurboQuant reduces KV‑Cache memory by up to 4.6×, speeds 3‑bit attention computation up to 8× on H100, and delivers near‑zero accuracy loss across long‑context benchmarks, with open‑source implementations for Metal, vLLM and llama.cpp.
The DYSCO (Dynamic Attention‑Scaling Decoding) algorithm, introduced by Princeton’s Chen Danqi team and NYU, eliminates the need for fine‑tuning and restores performance on long‑context tasks, delivering up to a 25% relative gain on 128K token benchmarks while adding only about 3.8% extra FLOPs.
The article analyzes the storage‑bandwidth bottleneck of agent‑style large language models, explains why traditional pre‑fill and decode architectures underutilize network resources, and details a dual‑channel loading and smart scheduling design that unlocks idle bandwidth, achieving up to 1.9× higher throughput in both offline and online inference workloads.
This article analyzes why multimodal large language models often fail to detect small objects, identifies three core bottlenecks, and presents a four‑tiered optimization roadmap—from zero‑cost inference tricks to data augmentation, model fine‑tuning, and engineering safeguards—backed by three real‑world case studies and actionable guidelines.
Alibaba’s Qwen3.5-397B-A17B, a 397‑billion‑parameter open‑source multimodal LLM, combines mixed linear attention with a sparse MoE architecture to achieve 8.6‑19× higher decoding throughput than Qwen3‑Max, supports 201 languages, and can be deployed via vLLM, Docker, Transformers, or SGLang with various optimization presets.
This article examines the evolution of re‑ranking systems from traditional pointwise models to a two‑stage generation‑evaluation framework, compares autoregressive and non‑autoregressive generative approaches, details inference speed optimizations with GPU and model‑server upgrades, and outlines a future end‑to‑end sequence generation architecture enhanced by reinforcement learning and contrastive learning.
This article introduces Render‑of‑Thought (RoT), a novel paradigm that compresses chain‑of‑thought reasoning into visual embeddings using frozen vision encoders, achieving 3‑4× token reduction, faster inference, and improved interpretability while requiring minimal pre‑training.
This article details the engineering challenges and solutions for deploying a 3.5 billion‑parameter MoE LLM in Taobao's search relevance pipeline, covering large‑batch scheduling, dynamic load balancing, intra‑batch KV‑Cache reuse, and MoE kernel tuning to meet sub‑second latency requirements.
This article explains how SGLang’s hybrid model design combines Transformer attention with Mamba state‑space layers, introduces a dual‑pool memory architecture and elastic allocation, and presents specialized prefix‑cache and speculative‑decoding techniques that together enable efficient, scalable inference for long‑context large language models.
After discovering that only a few vLLM settings truly impact performance, this guide details how adjusting gpu_memory_utilization, max_num_batched_tokens, and enabling chunked prefill can raise Qwen2.5‑72B‑Instruct throughput from ~1800 to over 2500 tokens/s, improve latency, and provides comprehensive deployment, monitoring, and troubleshooting instructions.
This article walks through the complete workflow of AI large language models, from turning user queries into token matrices via tokenization and embedding, through the Transformer’s self‑attention and multi‑head mechanisms, to decoding logits into human‑readable text, while also covering position encoding, long‑context strategies, generation parameters, and practical engineering tips.
SpecExit combines speculative sampling with a lightweight draft model to predict early‑exit signals, shortening large‑reasoning model chains by up to two‑thirds and achieving up to 2.5× end‑to‑end inference acceleration on vLLM without sacrificing accuracy.
Baidu Baige and the SGLang community have open‑sourced a production‑tested MTP inference engine that boosts DeepSeek‑V3.2 decoding speed by over two times while delivering exceptional stability, thanks to a DSA‑optimized architecture that predicts multiple tokens in a single forward pass.
The paper presents a systematic scaling‑law study of the linear‑time xLSTM architecture versus quadratic‑time Transformers, evaluating parameter‑data loss surfaces, optimal model size under equal FLOP budgets, and inference latency components, and shows that xLSTM consistently offers better cost‑effectiveness across diverse contexts and budgets.
Ant Group's open‑source dInfer framework dramatically speeds up diffusion language model inference—achieving more than a ten‑fold boost over Fast‑dLLM, surpassing autoregressive baselines, and delivering 1011 tokens per second on HumanEval—by tackling computational cost, KV‑cache invalidation, and parallel decoding challenges through modular system‑level innovations.
This article explains the EasyRec recommendation system’s training and inference architecture, detailing optimization techniques such as embedding parallelism, CPU/GPU placement, XLA and TRT fusion, online learning pipelines, network compression, and real‑world deployment results that dramatically improve throughput and latency.
DeepSeek‑V3.2‑Exp, the latest experimental large‑language model, is open‑sourced with a paper, featuring a million‑token context window, a new sparse attention mechanism, GRPO‑enhanced reasoning, and detailed cost‑analysis showing up to ten‑fold inference savings.
Ring-mini-2.0 is a high‑performance inference MoE model that activates only 1.4 B parameters out of 16 B total, achieving dense‑model quality below 10 B while supporting 128 K context length and ultra‑fast generation speeds of over 300 tokens/s.
This article explains why Meituan introduced a generative recommendation model, describes the MTGR architecture, data organization, training and inference engines built on TorchRec and TensorRT, reports performance gains and cost reductions, and outlines future directions such as simplifying the recommendation funnel and cross‑business heterogeneous modeling.
This article walks through installing and configuring vLLM for local large language model inference, compares it with Ollama and LM Studio, details environment setup, model download, testing scripts, and shows how to expose an OpenAI‑compatible API for production use.
Meituan has open‑sourced LongCat‑Flash‑Chat, a 5.6‑trillion‑parameter Mixture‑of‑Experts model that activates only a fraction of its weights per token, delivering mainstream‑level performance, high inference speed, and low cost for complex agent applications.
LongCat‑Flash is a 560‑billion‑parameter Mixture‑of‑Experts LLM that combines a dynamic zero‑computation expert design, shortcut‑connected MoE communication, variance‑aligned scaling, and a three‑stage agent‑centric pre‑training pipeline, delivering over 100 TPS on H800 GPUs at a cost of $0.70 per million tokens.
This article explains why pre‑training large AI models consumes 90‑99% of total compute, describes the full training and inference pipelines, introduces resource‑saving strategies such as PD‑separation, and reviews market trends and infrastructure challenges shaping the next generation of AI systems.
Lumina-mGPT 2.0 is a decoder‑only, zero‑shot trained autoregressive image model that rivals diffusion systems like DALL·E 3 in quality while offering unified multimodal tokenization, flexible multi‑task generation, and several inference‑speed tricks, yet it still faces licensing, scaling and sampling‑time challenges.
This article explores the evolution of AI infrastructure, comparing it with traditional backend systems, and details how hardware shifts to GPU-centric designs, software adaptations like deep learning frameworks, and engineering challenges in model training and inference can be addressed using established backend methodologies.
This article details the Angel‑HCF team's end‑to‑end DeepSeek inference optimizations—including PD separation, multi‑layer MTP, EP and DP parallelism, hardware‑aware kernels, and load‑balancing strategies—that boost throughput to over 15,800 tokens per second while keeping per‑token latency under 50 ms.
This article explains the EasyRec recommendation system's training and inference architecture, details a series of optimizations for both CPU and GPU pipelines, and describes the online learning workflow that enables real‑time model updates across large‑scale e‑commerce scenarios.
This article explains EasyRec's end‑to‑end recommendation system, covering its training‑inference architecture, a series of CPU/GPU and distributed optimizations, and a real‑time online‑learning pipeline that together improve throughput, latency, and model freshness.
This article traces the evolution of AI application architectures—from the earliest minimal user‑LLM interaction to advanced designs featuring context enhancement, input/output guardrails, intent routing, model gateways, caching strategies, agent capabilities, monitoring, and inference performance optimizations—providing practical insights and references for developers.
Meituan’s recommendation team introduced the MTGR framework, aligning traditional DLRM features with a unified HSTU‑based Transformer to explore scaling laws, delivering a 65‑fold FLOPs boost, 12% lower inference cost, and record gains in online CTR and order volume across its food‑delivery platform.
This guide explains how to use Baidu Baige's AI heterogeneous computing platform to deploy the eight‑model Qwen3 family—including dense and MoE variants—via a one‑click process, covering resource configuration, inference acceleration options, and post‑deployment service access.
This article dissects the major bottlenecks of large‑language‑model serving—prefill vs. decode, sparsity, memory bandwidth, KV‑cache growth—and walks through concrete engineering tricks such as paged attention, radix‑tree KV caches, compressed attention, speculative decoding, FlexGen weight scheduling, FastServe queuing, plus a runnable vLLM code snippet.
JD’s advertising platform replaces rule‑based recall with a generative large‑model pipeline that unifies e‑commerce knowledge, multimodal user intent, and semantic IDs across recall, coarse‑ranking, fine‑ranking and creative optimization, while meeting sub‑100 ms latency and sub‑¥1‑per‑million‑token cost through quantization, parallelism, caching, and joint generative‑discriminative inference, delivering double‑digit performance gains and paving the way for domain‑specific foundation models.
This article provides a comprehensive technical guide for privately deploying DeepSeek large language models, covering model and runtime parameter selection, hardware sizing calculations, software stack preparation, inference service setup, performance tuning, and security monitoring considerations.
Cache‑augmented generation (CAG) preloads documents into LLM context using KV caches to eliminate retrieval latency, offering faster inference for static knowledge bases, while RAG remains more flexible for dynamic or large corpora; this article compares their definitions, performance, implementation steps, and future prospects.
PaddlePaddle 3.0 introduces a full‑stack inference engine that supports FP8, INT8, and 4‑bit quantization for popular LLMs such as DeepSeek V3/R1, delivers up to 2× token throughput on a single H800 GPU, and provides detailed deployment scripts for single‑node and multi‑node setups, including MTP speculative decoding and SageAttention for long‑sequence acceleration.
QwQ-32B, Alibaba’s new inference‑optimized large language model built on the Qwen2.5 architecture, outperforms DeepSeek‑R1 across math reasoning, code generation, and safety benchmarks while requiring only 24 GB vRAM, and the article provides detailed performance data, resource‑efficiency analysis, and step‑by‑step Java and Ollama integration instructions.
The Inspur Metabrain R1 inference server, equipped with FP8 acceleration and a 1128 GB HBM3e memory pool, has been tightly integrated with SGLang 0.4.3 to run the 671‑billion‑parameter DeepSeek R1 model, delivering over 1,000 concurrent user sessions and up to 3,976 tokens/s throughput.
The article analyzes DeepSeek’s recent breakthroughs—including its Mixture‑of‑Experts architecture, cost‑effective inference optimizations, high‑accuracy multimodal processing, and open‑source collaboration—while also offering a curated bundle of technical e‑books covering AI chips, networking, storage, and more.
The article describes an edge‑based video‑cover recommendation system for DeWu that leverages the MNN SDK and a lightweight MobileNetV3 model, performing on‑device inference with quantization and parallel processing to automatically select high‑quality covers, achieving sub‑second latency and boosting click‑through rates by up to 18 %.
JD Advertising accelerates its generative‑recall recommendation system by integrating NVIDIA TensorRT‑LLM, which simplifies the pipeline, injects LLM knowledge, scales to billions of parameters, and delivers over five‑fold throughput gains, one‑fifth the cost, and significant CTR improvements in both recommendation and search.
Since the launch of ChatGPT, 58.com has built a Model‑as‑a‑Service platform called LingXi that trains and serves domain‑specific large language models, supports over a hundred internal scenarios with daily inference exceeding ten million calls, and continuously improves performance through quantization, GPU optimization, model miniaturization, and advanced AI applications such as interview assistants, voice agents, and RAG‑enabled agents.
This article explains how generative recommendation systems powered by large language models simplify the recommendation pipeline, integrate world knowledge, benefit from scaling laws, and require specialized engineering optimizations such as TensorRT‑LLM deployment, inference acceleration, and hybrid model strategies to achieve low latency and high throughput in real‑world e‑commerce scenarios.
The article examines DeepSeek‑V3’s architecture and training pipeline, highlighting its use of MLA and a highly granular MoE design, pioneering FP8 mixed‑precision training, fine‑grained per‑tile quantization, advanced parallelism strategies, and inference optimizations such as PD separation and NanoFlow to achieve unprecedented efficiency on limited GPU resources.
This article explains how Momo utilizes large language model technologies to revamp its AI application paradigm, achieve efficient inference through quantization and prefix caching, build a workflow‑based model platform, and outline future plans for framework optimization and multimodal support.
This article presents a comprehensive overview of EasyRec’s recommendation system architecture, detailing training and inference optimizations, embedding parallelism, CPU/GPU placement strategies, online learning pipelines, and network compression techniques that together improve scalability, latency, and cost efficiency.
At the 2024 Yunqi Conference, Alibaba Cloud’s AI Infra experts detailed the latest challenges of large‑model deployment—such as hardware costs, resource management, and software‑hardware coordination—and introduced PAI’s new capabilities, including stability tools, automated distributed training, reinforcement‑learning frameworks, inference optimizations, and integrated big‑data AI solutions.
EasyRec, Alibaba Cloud’s modular recommendation framework, unifies configurable data, embedding, dense, and output layers on MaxCompute, EMR, and DLC, and speeds training with deduplication, EmbeddingParallel sharding, lock‑free hash tables, GPU embeddings, and AMX BF16, while inference benefits from operator fusion, low‑precision AVX/AMX kernels, compact caches, batch merging, and network compression, enabling real‑time online learning and delivering higher recommendation quality at lower cost in e‑commerce.
This article presents a comprehensive overview of EasyRec's recommendation system architecture, detailing training and inference optimizations, distributed deployment strategies, operator fusion techniques, online learning pipelines, and network-level improvements to enhance performance and scalability.
This guide presents a hands‑on curriculum of core large‑model engineering tasks—including model conversion scripts, custom modeling wrappers, multi‑model inference utilities, and channel‑aware loss tracking—to help practitioners build practical, reusable tools without deep theoretical overhead.
Baidu Baige’s AIAK‑LLM suite accelerates large‑model training and inference by boosting Model FLOPS Utilization through techniques such as TP communication overlap, hybrid recompute, zero‑offload, automatic parallel‑strategy search, multi‑chip support, and inference‑specific optimizations, achieving over 60 % speedup and seamless Hugging Face integration.
The talk outlines how Baidu’s Baige AIAK‑LLM suite tackles the exploding compute demands of trillion‑parameter models by boosting Model FLOPS Utilization through advanced parallelism, memory‑saving recompute, zero‑offload, adaptive scheduling, and cross‑chip orchestration, delivering 30‑60% training and inference speedups and a unified cloud product.
By fusing sparse‑feature operators, enabling multi‑stream execution, consolidating data copies, and merging inference batches, iQIYI reduced GPU CTR model latency to CPU‑level, boosted throughput over sixfold, and cut operational costs by more than 40%, overcoming launch‑overhead bottlenecks.
This article explains two production‑grade optimization techniques for large language model inference—KV‑cache reuse across multi‑turn dialogues and speculative sampling with a small draft model—detailing their design, implementation, and performance impact.
This article explains how to use PAI-Blade to accelerate Stable Diffusion inference by optimizing LoRA and ControlNet components, detailing configuration steps, code modifications, benchmark results on A100/A10 GPUs, and integration with both Diffusers and the popular Stable-Diffusion-webui, highlighting performance gains and memory savings.
This article explains how PAI‑Blade, built on the BladeDISC compiler and BlaDNN library, dramatically reduces latency and memory usage for Stable Diffusion inference, provides step‑by‑step usage examples with code, shows performance gains on A100 and A10 GPUs, and outlines future optimization directions.
The article presents high_service, a high‑performance inference framework that boosts GPU utilization in multimedia AI services by separating CPU‑heavy preprocessing from GPU inference, employing priority‑based auto‑scaling, multi‑tenant sharing, and TensorRT‑accelerated models to eliminate GIL bottlenecks, reduce waste, and adapt to fluctuating traffic, with future work targeting automated bottleneck detection and further CPU‑GPU offloading.
This article reviews DeepSpeed‑MoE, an end‑to‑end system that introduces new MoE architectures, model‑compression techniques, and highly optimized inference pipelines, detailing its motivation, design of PR‑MoE (Pyramid‑MoE and Residual‑MoE), distributed parallel strategies, communication and kernel optimizations, and performance gains over dense baselines.
With the rapid rise of edge computing, deploying AI models for tasks like object detection, OCR, and speech recognition on resource‑constrained devices faces speed challenges; the upgraded Paddle Lite inference engine and PaddleSlim compression tools claim up to 23% faster inference and significant model size reductions, offering a practical solution.
This article reviews the evolution of pretraining models in NLP, discusses the practical challenges of deploying large models such as inference latency, knowledge integration, and task adaptation, and presents Xiaomi’s optimization techniques including knowledge distillation, low‑precision inference, operator fusion, and multi‑granularity segmentation for dialogue systems.
The article details the development, architecture, and optimization of 58.com’s AI algorithm platform (WPAI), covering its background, overall design, large‑scale distributed machine learning, deep‑learning platform features, inference performance enhancements, GPU resource scheduling improvements, and future directions.
This article shares the Xiaomi AI Lab NLP team's experience in the NLPCC lightweight language model competition, discusses efficiency challenges of large pretrained models like BERT, and details practical inference optimizations—including model distillation, batching, FP16 quantization, and FasterTransformer integration—that dramatically reduce latency and hardware costs in production.
To alleviate GPU scarcity, iQIYI’s cloud platform migrated deep‑learning inference to CPUs and applied system‑level (MKL‑DNN, OpenVINO), application‑level, and algorithm‑level optimizations—tuning threads, batch size, NUMA, pruning and quantization—delivering 1‑9× speedups across thousands of cores while preserving latency and accuracy.
This article explains how Alipay’s xNN engine overcomes mobile deep‑learning challenges through aggressive model compression, lightweight SDK design, algorithm‑ and instruction‑level optimizations, enabling high‑accuracy AI inference on a wide range of Android and iOS devices with minimal app‑size impact.
